Method for small sample radiation source individual identification based on top-k neighborhood and prototype learning

Abstract

This invention belongs to the field of machine learning and signal processing technology, specifically relating to a method for identifying individual radiation sources in small samples based on Top-K neighborhood and prototype learning. This method constructs an online network and a target network, combining strong and weak data augmentation, exponential moving average updates, and a dynamic feature queue to obtain historical key vectors; it retrieves Top-K local neighborhoods using normalized cosine similarity, and constructs positive and negative sample sets within the neighborhoods based on label information to calculate local contrast loss; simultaneously, it constructs category prototypes based on the mean of key vectors of the same category and calculates prototype loss; during the testing phase, it performs similarity discrimination based on the category prototypes constructed during the training phase. This invention's method can learn stable and discriminative radiation source fingerprint features under conditions of limited sample size, sparse feature distribution, and fluctuating intra-class variance, improving the accuracy and generalization robustness of identifying individual radiation sources in small samples.

Description

Technical Field

[0001] This invention belongs to the field of machine learning and signal processing technology, specifically relating to a method for identifying individual radiation sources in small samples based on Top-K neighborhood and prototype learning. Background Technology

[0002] Existing research on radiation source identification typically implicitly assumes a sufficient number of training samples to ensure that deep models can learn stable fingerprint features from large-scale data. However, in real-world electromagnetic environments, limitations such as acquisition conditions, scarcity of equipment types, and high costs of manual annotation often result in only a very small number (several to dozens per class) of labeled radiation source signal samples. This scarcity of training data creates a fundamental contradiction with the implicit dependence of existing methods on data volume, making small-sample learning a critical technical bottleneck that urgently needs to be overcome in this field.

[0003] Few-shot learning aims to train a model using only a small number of labeled samples, enabling it to generalize well to unknown categories. However, in the task of identifying individual radiation sources, the extremely low sample size leads to two core challenges for existing methods: First, overfitting—deep models tend to "memorize" the training data rather than generalize features on a limited number of samples, resulting in significant performance degradation on the test set; second, discriminative feature learning is difficult—under very few samples, the model struggles to extract stable, inter-class separable, and intra-class compact radiation source fingerprint features, severely limiting the improvement of recognition accuracy.

[0004] To address the aforementioned issues, there is an urgent need for a self-supervised representation learning method that can fully exploit local structural information in limited samples under low-sample conditions and explicitly introduce category semantic constraints, in order to improve the accuracy and generalization ability of identifying individual radiation sources in small samples. Summary of the Invention

[0005] The present invention aims to address the aforementioned problems by proposing a radiation source individual identification method that combines momentum comparison and self-distillation mechanism, is suitable for low sample conditions, and has a certain robustness to sparse feature distribution and intra-class variance fluctuations, i.e., has a higher identification accuracy.

[0006] The technical solution of this invention is: a radiation source individual identification method that combines momentum contrast and self-distillation mechanisms, is suitable for conditions with very few samples, and has a certain robustness to sparse feature distribution and intra-class variance fluctuations, i.e., it has higher identification accuracy. Its main feature is the introduction of a dynamic feature queue within the online network and target network framework. Normalized cosine similarity is used to retrieve Top-K local neighborhoods from the dynamic feature queue, and positive and negative sample sets are constructed in the neighborhoods based on label information to calculate local contrast loss. Simultaneously, a class prototype is constructed based on the mean of the same-class key vectors, and prototype loss is calculated, so that the training process is constrained by both the local neighborhood structure and the global class structure. This method enables the system to effectively alleviate overfitting problems under conditions of very few samples, sparse feature distribution, and intra-class variance fluctuations, extracting stable and discriminative radiation source fingerprint features, and achieving higher accuracy in small-sample radiation source individual identification tasks. The implementation process includes the following steps:

[0007] The technical solution of this invention is as follows:

[0008] A method for identifying individual radiation sources based on Top-K neighborhood and prototype learning in small samples includes the following steps:

[0009] S1. Obtain a small sample training set of radiation source signals. ,in This is the preprocessed I / Q sequence. Used to represent samples The individual category label of the radiation source to which it belongs. It is the number of categories. That is the number of samples in each category;

[0010] S2. Construct an online network and a target network, wherein the online network includes an encoder, a projection layer and a prediction layer, and the target network includes an encoder and a projection layer;

[0011] Based on training samples Construct weakly enhanced samples respectively and strongly enhanced samples Weakly enhanced samples Input the target network to extract features and obtain the key vector. and their corresponding tags , will enhance the sample The input is used for feature extraction in an online network to obtain a value vector. :

[0012] ,

[0013] ,

[0014] Among them, among them, , These represent the projection layer and encoder of the target network, respectively. , , These represent the prediction layer, projection layer, and encoder of the online network, respectively.

[0015] S3. Train the constructed online network and the target network. The online network is updated through gradient backpropagation, while the target network is updated using an exponential moving average based on the parameters of the online network.

[0016] ,

[0017] in, For the target network parameters, For online network parameters, The momentum coefficient;

[0018] The specific training process is as follows:

[0019] For historical sample data, a dynamic feature queue is constructed. Key vectors used to store historical samples extracted by the target network and their corresponding tags , It is the number of historical samples and is updated in a first-in-first-out manner. When a new key vector and its corresponding label enter the queue, the earliest key vector and its corresponding label stored in the queue are removed.

[0020] Calculate the key vector of the current sample extracted by the target network. Key vectors in the dynamic feature queue The normalized cosine similarity between them is calculated using the following formula:

[0021] ,

[0022] And according to similarity Selection and The top-K most similar samples constitute the local neighborhood of the current sample:

[0023] ,

[0024] in, To surround the sample Constructed Top-K local neighborhoods;

[0025] In a batch, for all constructed neighborhoods Based on the current sample label and the labels of each sample in the local neighborhood, a positive sample set and a negative sample set are constructed, where the positive sample set and the negative sample set are represented as follows:

[0026] ,

[0027] ,

[0028] in, For the positive sample set, The set of negative samples; when there are filled samples or invalid labeled samples during the local neighborhood construction process, the filled samples or invalid labeled samples are removed from the positive sample set and the negative sample set;

[0029] Based on the positive sample set and negative sample set The local contrast loss is calculated using the following formula:

[0030] ,

[0031] in, For temperature parameters;

[0032] The category prototype is calculated based on the key vectors of the same category in the dynamic feature queue, where the first... The class prototype is represented as follows:

[0033] ,

[0034] in, For the first Class prototypes; and normalize the class prototypes;

[0035] Based on the value vector output by the online network and corresponding category prototypes The prototype loss is calculated using the following formula:

[0036] ,

[0037] in, The prototype loss temperature parameter is used to constrain the online network output features to align with the category prototype corresponding to its label.

[0038] The total loss is obtained by weighting and combining the local contrast loss and the prototype loss, and the online network parameters are updated based on the total loss. The calculation formula is as follows:

[0039] ,

[0040] in, This indicates the samples contained in a batch. This represents the number of samples in a batch. The prototype loss weights are used; when the number of valid samples in the dynamic feature queue does not meet the local neighborhood construction condition, the dynamic feature queue is updated; when the local neighborhood construction condition is met, the online network parameters are updated based on the local contrast loss, and when the prototype constraint activation condition is met, the online network parameters are updated together based on the local contrast loss and the prototype loss; after the set training conditions are met, the trained online network is obtained.

[0041] S4. Obtain a small sample of the radiation source target signal, and then... The target sample feature representation is obtained by inputting the trained online network. ;calculate The similarity between the target signal and the prototype of each category is used, and the category corresponding to the prototype of the category with the highest similarity is used as the radiation source individual identification result of the target signal.

[0042] The beneficial effects of this invention are as follows: By constructing Top-K local neighborhoods for comparative learning, the local manifold structure of the feature space in a limited number of samples can be fully explored, capturing subtle fingerprint differences between samples, thereby improving the discriminative power of features. Simultaneously, the introduction of class prototype constraints aligns features of similar samples towards the global class center, forming a feature distribution that is separable between classes and compact within classes, effectively alleviating the problems of model overfitting and feature learning difficulties under extremely low sample conditions. Based on this, this invention adopts a classification strategy based on class prototypes, eliminating the need for an additional parameterized classifier, and directly relying on the class structure already formed in the feature space for inference and decision-making, thus reducing model complexity and overfitting risk to a certain extent. This invention does not rely on large-scale labeled data or make prior assumptions about sample distribution, exhibiting higher recognition accuracy and generalization robustness in small-sample radiation source individual identification tasks. Attached Figure Description

[0043] Figure 1 This is the overall flowchart of the present invention.

[0044] Figure 2 It is the target network and online network graph for training.

[0045] Figure 3 This is a comparison chart between the present invention and a baseline that only uses network training.

[0046] Figure 4 This is a schematic diagram of the recognition results of the method of the present invention under a small sample setting.

[0047] Figure 5 This is a schematic diagram of the recognition results of the method of the present invention under another small sample setting. Detailed Implementation

[0048] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0049] The method of the present invention will now be described in detail through the training and testing process:

[0050] Suppose a small sample radiation source signal dataset This represents a dataset of radiation sources containing N categories, with K samples in each category.

[0051] During training, for each input sample Two different strengths of data augmentation were constructed: weak augmentation. (Such as amplitude variation, Gaussian white noise) and strong enhancement (e.g., time-frequency domain transformation, composite signal enhancement). Simultaneously, two coding networks with the same structure but different parameter update methods are employed—an online network. With the target network The online network is updated via gradient backpropagation, while the target network is updated via exponential moving average (EMA).

[0052]

[0053] in, This is the momentum coefficient.

[0054] Both networks employ an asymmetric data augmentation strategy: weakly augmented samples are input into the target network, while strongly augmented samples are input into the online network.

[0055] To fully utilize historical sample information, a dynamically updated feature queue is maintained. ,in The key vector extracted from the target network. For the corresponding labels. This queue uses a first-in, first-out (FIFO) strategy for updates, controlling storage size while ensuring sample diversity, and ensuring that key vectors are always generated by the latest network. For input samples Its key vector is calculated by the target network, and its value vector is... It is then calculated by an online network.

[0056]

[0057]

[0058] Among them, among them, , These represent the projection layer and encoder of the target network, respectively. , , These represent the prediction layer, projection layer, and encoder of the online network, respectively. This is used to measure the key vector of a sample. key vectors in the queue The similarity between them was determined using normalized cosine similarity:

[0059]

[0060] Based on this similarity, select from the queue... Construct the local neighborhood of the top-K most similar samples:

[0061]

[0062] This neighborhood can be considered as surrounding the sample. A local feature subspace is constructed, and samples in the neighborhood are further divided into positive and negative sample sets based on whether their labels match the current sample's label. In a batch, for all constructed neighborhoods... A set of positive and negative samples is constructed using label information. For each sample... The obtained key vector Its positive and negative sample set is defined as:

[0063]

[0064]

[0065] In actual training, because neighborhood construction may involve padding operations (when the number of available samples is insufficient at the start of training), some sample labels may be marked as... To avoid introducing noise, these invalid samples are removed from the neighborhood. Based on the above definition, samples... The constructed contrastive learning objectives are as follows:

[0066]

[0067] in, The temperature parameter controls the distribution sharpening. This contrastive loss function encourages samples to be closer to their neighbors of the same class in the feature space, while maintaining distinction from samples of different classes.

[0068] To characterize semantic information at the global level, a category prototype mechanism is further introduced. For each category... Its prototype is defined as the mean of that category feature in the queue:

[0069]

[0070] Online network output is After normalizing the prototype, the prototype loss is defined as follows:

[0071]

[0072] This prototype loss constrains features at the class level, causing them to cluster towards the corresponding class centers, thereby enhancing the global structure of the feature space. The final training loss function consists of both local contrastive loss and prototype loss.

[0073]

[0074] in, A batch contains the samples. , This represents the number of samples in a batch. The overall loss function achieves multi-level optimization of the feature space by simultaneously constraining local consistency and global semantic structure.

[0075] During the testing phase, a category prototype-based classification strategy is used to predict test samples. Unlike the training phase, this phase only uses the already trained online network and does not perform any parameter updates, thus ensuring the efficiency and stability of the inference process. Simultaneously, the model directly utilizes the category prototypes constructed during the training phase to classify the test samples. Given a test sample... First, through a pre-trained online network Extract its feature representation:

[0076]

[0077] The category prototype is the category prototype with the minimum loss that is saved during the training phase. For each category... Its prototype is defined as all categories in the queue. key vector The mean of all prototypes. All prototypes constitute the prototype matrix:

[0078]

[0079] in, This refers to the dimension of the key vector. For test features... Calculate its similarity to the prototypes of each category:

[0080]

[0081] The final predicted category is obtained using the maximum similarity principle:

[0082]

[0083] Example:

[0084] This example demonstrates the identification of small-sample radiation sources using the method proposed in this invention. The overall flowchart is as follows: Figure 1 As shown:

[0085] The specific steps of the contrastive learning training algorithm based on Top-K neighborhood and prototype learning are as follows:

[0086] Step 1: Obtain a small sample training set of radiation source signals ,in This is the preprocessed I / Q sequence of the radiation source signal. For sample labels.

[0087] Step 2: Pre-trained encoder network One cycle.

[0088] Step 3: Copy the pre-trained encoder parameters to both the online network and the target network, such as... Figure 2 As shown.

[0089] Step 4: Randomly perform strong and weak data augmentation on the input samples, with the strong augmentation sample... Send to online network to get Weakly enhanced samples Send to the target network to obtain Weak enhancement is achieved by adding 25–30 dB of random Gaussian white noise to the signal; strong enhancement also adds 20–25 dB of Gaussian white noise and further applies a time mask with a mask ratio of 0.1. At the same time, the strong enhancement sample will have a 50% probability of performing a signal flipping operation.

[0090] Step 5: When the training period is less than At this time, only queue filling is performed, and network parameter updates are not performed. .

[0091] Step 6: Calculate the key vector of the sample With the queue Similarity score, and select the top-K most similar pairs. Forming a neighborhood .

[0092] Step 7: Combine all neighborhoods to construct positive and negative sample neighborhoods.

[0093] Step 8: Calculate the category prototype and L2 normalize it.

[0094] Step 9: When the training period is less than At that time, only the contrast loss is calculated. Update online network parameters. .

[0095] Step 10: When the training period is greater than At that time, the contrast loss and prototype loss are calculated, and the online network parameters are updated.

[0096] Step 11: EMA updates comparison network parameters.

[0097] Step 12: Update the queue, removing the oldest key vector to keep the queue size constant.

[0098] The specific steps of the comparative learning testing algorithm based on Top-K neighborhood and prototype learning are as follows:

[0099] Step 1: Obtain a small sample test set of radiation source signals ,in This is the preprocessed I / Q sequence of the radiation source signal. For sample labels.

[0100] Step 2: The test samples are directly fed into the online network to obtain feature representations. .

[0101] Step 3: Calculation Similarity to the category prototype.

[0102] Step 4: Select the category represented by the prototype with the highest similarity as the test result. Figures 3 to 5 This is used to illustrate the recognition results of this embodiment under different settings.

Claims

1. A method for identifying individual radiation sources from a few samples based on Top-K neighborhood and prototype learning, characterized in that, Includes the following steps: S1. Obtain a small sample training set of radiation source signals. ,in This is the preprocessed I / Q sequence. Used to represent samples The individual category label of the radiation source to which it belongs. It is the number of categories. That is the number of samples in each category; S2. Construct an online network and a target network, wherein the online network includes an encoder, a projection layer and a prediction layer, and the target network includes an encoder and a projection layer; Based on training samples Construct weakly enhanced samples respectively and strongly enhanced samples Weakly enhanced samples Input the target network to extract features and obtain the key vector. and their corresponding tags , will enhance the sample Inputting the data into an online network for feature extraction yields a value vector. : ， ， Among them, among them, , These represent the projection layer and encoder of the target network, respectively. , , These represent the prediction layer, projection layer, and encoder of the online network, respectively. S3. Train the constructed online network and the target network. The online network is updated through gradient backpropagation, while the target network is updated using an exponential moving average based on the parameters of the online network. ， in, For the target network parameters, For online network parameters, The momentum coefficient; The specific training process is as follows: For historical sample data, a dynamic feature queue is constructed. Key vectors used to store historical samples extracted by the target network and their corresponding tags , It is the number of historical samples and is updated in a first-in-first-out manner. When a new key vector and its corresponding label enter the queue, the earliest key vector and its corresponding label stored in the queue are removed. Calculate the key vector of the current sample extracted by the target network. Key vectors in the dynamic feature queue The normalized cosine similarity between them is calculated using the following formula: ， And according to similarity Selection and The top-K most similar samples constitute the local neighborhood of the current sample: ， in, To surround the sample Constructed Top-K local neighborhoods; In a batch, for all constructed neighborhoods Based on the current sample label and the labels of each sample in the local neighborhood, a positive sample set and a negative sample set are constructed, where the positive sample set and the negative sample set are represented as follows: ， ， in, For the positive sample set, The set of negative samples; when there are filled samples or invalid labeled samples during the local neighborhood construction process, the filled samples or invalid labeled samples are removed from the positive sample set and the negative sample set; Based on the positive sample set and negative sample set The local contrast loss is calculated using the following formula: ， in, For temperature parameters; The category prototype is calculated based on the key vectors of the same category in the dynamic feature queue, where the first... The class prototype is represented as follows: ， in, For the first Class prototypes; and normalize the class prototypes; Based on the value vector output by the online network and corresponding category prototypes The prototype loss is calculated using the following formula: ， in, The prototype loss temperature parameter is used to constrain the online network output features to align with the category prototype corresponding to its label. The total loss is obtained by weighting and combining the local contrast loss and the prototype loss, and the online network parameters are updated based on the total loss. The calculation formula is as follows: ， in, This indicates the samples contained in a batch. This represents the number of samples in a batch. The prototype loss weights are used; when the number of valid samples in the dynamic feature queue does not meet the local neighborhood construction condition, the dynamic feature queue is updated; when the local neighborhood construction condition is met, the online network parameters are updated based on the local contrast loss, and when the prototype constraint activation condition is met, the online network parameters are updated together based on the local contrast loss and the prototype loss; after the set training conditions are met, the trained online network is obtained. S4. Obtain a small sample of the radiation source target signal, and then... The target sample feature representation is obtained by inputting the trained online network. ;calculate The similarity between the target signal and the prototype of each category is used, and the category corresponding to the prototype of the category with the highest similarity is used as the radiation source individual identification result of the target signal.

Images