A small sample radiation source individual open set identification method

Abstract

The application belongs to the technical field of signal processing, and discloses a small sample radiation source individual open set identification method, first, a deep fusion network composed of a CNN and a pre-training Transformer architecture semantic model is constructed, and a progressive fine-tuning strategy is used to realize small sample transfer learning of the radiation source. Meanwhile, fingerprint features with clear physical meaning are extracted from the radiation source IQ signal, and the distribution of the known categories is modeled based on the Mahalanobis distance. In the identification stage, the classification confidence of the deep fusion network and the Mahalanobis distance model confidence state and distance ratio based on the physical fingerprint features are obtained in parallel, and through the designed multi-scene fusion decision rule, the sample belonging to the known category or the unknown category is judged in cooperation. The application can realize high-precision closed set classification under a small amount of labeled samples, and effectively detect unknown radiation source individuals in the open set, and significantly improve the practicability and reliability of the radiation source individual identification in the complex electromagnetic environment.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, and in particular relates to a method for identifying individual open sets of small-sample radiation sources. Background Technology

[0002] Individual radiation source identification and refined analysis are key tasks in cognitive electronic warfare. Faced with increasingly complex electromagnetic backgrounds and non-cooperative, highly adversarial battlefield environments, traditional methods for individual radiation source identification lack reliability and cannot meet real-world demands. While artificial intelligence methods such as deep learning have achieved some success, their reliance on large datasets contradicts the reality of small sample sizes in engineering practice. Furthermore, previous radiation source research was entirely based on the closed-set assumption: that all radiation source types to be identified have appeared in the training samples. In actual classification tasks, especially in non-cooperative scenarios, sample collection becomes extremely difficult. In most cases, the collected sample set will fall far short of covering all radiation source categories, rendering the closed-set assumption invalid. Therefore, open-set identification techniques for individual radiation sources under small sample conditions not only require proper data processing but also the ability to correctly identify known categories while rejecting unknown samples, making them even more significant for research.

[0003] The existing methods have the following shortcomings:

[0004] 1. Traditional methods for identifying individual radiation sources lack reliability.

[0005] In the face of complex electromagnetic backgrounds and non-cooperative, highly antagonistic environments, traditional methods are insufficient to meet practical needs.

[0006] 2. Deep learning methods are poorly adapted to small sample sizes.

[0007] Deep learning relies on large amounts of labeled data, which contradicts the "small sample" conditions in real-world engineering.

[0008] 3. The closed-set assumption does not fit the actual application scenario.

[0009] Most existing studies are based on the "closed set hypothesis," which states that the training samples cover all categories to be identified. However, in non-cooperative scenarios, sample collection is difficult and it is hard to cover all radiation source categories, so the closed set hypothesis does not hold.

[0010] 4. Purely data-driven methods have weak discrimination ability for unknown samples.

[0011] When faced with categories that have never appeared before (out-of-distribution samples), there is a lack of clear and stable decision-making criteria, and the ability to identify open sets is limited.

[0012] 5. Lack of a fusion mechanism combining semantic features and physical fingerprints

[0013] Existing methods fail to effectively combine semantic features extracted by deep learning with fingerprint features that have clear physical meaning, thus limiting the reliability and interpretability of open set recognition. Summary of the Invention

[0014] The purpose of this invention is to provide a method for identifying individual open sets of radiation sources in a small sample, so as to solve the above-mentioned technical problems.

[0015] To solve the above-mentioned technical problems, the specific technical solution of the small-sample open-set identification method for radiation sources of the present invention is as follows:

[0016] A method for identifying individual open sets of radiation sources in a small sample, comprising the following steps:

[0017] Step 1: Obtain the complex baseband IQ signals of different radiation source individuals, and construct the training set, validation set, and test set for closed-set identification, and the test set for open-set identification;

[0018] Step 2: Construct a deep fusion network and train it using a progressive "freeze-fine-tuning" strategy;

[0019] Step 3: Extract physical layer fingerprint features from the IQ signal and construct a Mahalanobis distance model based on known class samples to set the classification threshold;

[0020] Step 4: Input the radiation source IQ signal into the trained deep fusion network, and simultaneously input the physical fingerprint extracted from the radiation source IQ signal into the Mahalanobis distance model to obtain the classification confidence of the deep learning neural network and the confidence state and distance ratio of the Mahalanobis distance criterion in parallel.

[0021] Step 5: Based on the combined scenario of deep fusion network confidence and Mahalanobis distance criterion, perform refined fusion decision-making and output the final open set recognition result.

[0022] Furthermore, the deep fusion network described in step 2 includes: a CNN front-end processing module, a projection module, a BERT module, and a classification head module. The CNN front-end processing module is used to extract local spatiotemporal features from the IQ signal, and this part consists of multiple layers of one-dimensional convolution, batch normalization, ReLU activation, and pooling layers. The projection module is used for data dimension matching. The BERT module is a semantic pre-trained model based on the Transformer architecture. The BERT module utilizes the powerful sequence modeling capabilities of the pre-trained BERT model to further enhance the contextual representation of features. The classification head module consists of an adaptive pooling layer and multiple fully connected layers, outputting the confidence score for each known category.

[0023] Furthermore, the gradual strategy of "freeze-fine-tune" described in step 2 includes two stages:

[0024] Phase 1: Freeze all values ​​of the BERT modules and train only the CNN front-end processing module, projection module, and classification head module. The training continues until the model's recognition accuracy on the closed-set validation set exceeds a preset threshold. Training terminates when the threshold is not reached. Then proceed to stage two;

[0025] Phase Two: Unfreeze the weights of the last two layers of the BERT module and fine-tune the entire network using a hierarchical learning rate strategy to enable the model to learn more refined features.

[0026] Furthermore, the physical layer fingerprint features described in step 3 include IQ imbalance, carrier frequency offset, phase noise features, amplitude distortion, inter-symbol interference features, instantaneous phase features, and instantaneous amplitude features.

[0027] Furthermore, the Mahalanobis distance model construction in step 3 includes standardizing the physical layer fingerprint features of known category samples, calculating the mean vector and covariance matrix of each known category, performing regularization on the covariance matrix, and finally setting a Mahalanobis distance threshold for each category based on the chi-square distribution.

[0028] Furthermore, the deep learning classification confidence in step 4 is the maximum probability value after the output layer of the deep fusion network is processed by the Softmax function; the Mahalanobis distance decision result is: calculate the Mahalanobis distance from the physical layer fingerprint feature of the test sample to each known category. If the minimum distance is less than the threshold of the corresponding category, it is determined to be the corresponding known category; otherwise, it is determined to be the unknown category; the distance ratio is the ratio of the minimum Mahalanobis distance to the threshold of the corresponding category.

[0029] Furthermore, the fusion decision in step 5 is as follows:

[0030] a) If the confidence score of the deep learning classification is greater than the first threshold, and the Mahalanobis distance decision result is a known category, then the final decision is the known category corresponding to the deep learning classification result; b) If the confidence score of the deep learning classification is greater than the first threshold, and the Mahalanobis distance decision result is an unknown category, then the distance ratio is further determined: if the distance ratio is less than the second threshold, then the final decision is the known category corresponding to the deep learning classification result; otherwise, the final decision is the unknown category; c) If the confidence score of the deep learning classification is less than the first threshold, and the Mahalanobis distance decision result is a known category, then the final decision is the known category corresponding to the Mahalanobis distance decision result; d) If the confidence score of the deep learning classification is less than the first threshold, and the Mahalanobis distance decision result is an unknown category, then the final decision is the unknown category.

[0031] The small-sample open-set identification method for individual radiation sources of the present invention has the following advantages:

[0032] High-precision closed-set classification: By fusing CNN with a pre-trained BERT semantic model and employing a progressive training strategy, high-precision differentiation between known classes can still be achieved under small sample conditions; Stable and reliable open-set recognition: Mahalanobis distance, based on the physical layer features of the signal, is more sensitive to changes in unknown classes and exhibits high stability; Intelligent decision fusion: The proposed fusion strategy effectively combines the closed-set classification capability of deep learning with the out-of-distribution open-set detection capability of Mahalanobis distance, cross-validating each other and significantly reducing the probability of misclassification of known classes and omission of unknown classes; Strong engineering practicality: The entire system does not require a complex model training process, has a clear workflow, and is easy to implement and deploy. Attached Figure Description

[0033] Figure 1 This is a flowchart of the present invention;

[0034] Figure 2 This is a system block diagram of the present invention;

[0035] Figure 3 This is a visualization diagram of the T-SNE deep fusion network of the present invention;

[0036] Figure 4 A visualization of the T-SNE of the ResNet 1D network;

[0037] Figure 5 A graph showing the overall recognition rate of open set recognition algorithms;

[0038] Figure 6 This is a schematic diagram of the open set recognition confusion matrix when the openness is 0.106 in this invention. Detailed Implementation

[0039] To better understand the purpose, structure, and function of this invention, the following detailed description of a method for identifying individual open sets of small-sample radiation sources is provided in conjunction with the accompanying drawings.

[0040] In recent years, pre-trained large-scale model technology based on the Transformer architecture has provided a new approach to solving the small-sample problem. The fundamental knowledge and powerful contextual learning capabilities gained from pre-training on massive amounts of unlabeled data can be efficiently adapted to downstream tasks through transfer learning. Introducing it into the field of radiation source identification can effectively overcome the dependence of traditional deep learning on large-scale labeled signals, achieving stable transfer from "small samples" to "high accuracy." Furthermore, its core self-attention mechanism enables global perception and dependency modeling of signal sequences, thereby capturing long-distance temporal features and deep semantic information that are difficult for traditional convolutional networks to learn, providing more robust feature representations for individual identification.

[0041] Meanwhile, hardware defects inherent in radiation source signals, such as power amplifier nonlinearity and crystal oscillator deviation, manifest as unique physical fingerprint characteristics (e.g., phase noise, frequency shift, IQ imbalance). These characteristics are strongly correlated with specific individuals and exhibit distinguishable distribution patterns across different categories, making them crucial features for individual radiation source identification. These physical fingerprint characteristics possess clear statistical physical meaning and distributional properties, are insensitive to changes in model structure, and exhibit high stability. Therefore, in open-set identification tasks, physical fingerprints provide a stable and interpretable metric for defining "known" and "unknown" samples, effectively compensating for the inherent deficiency of purely data-driven methods in making decisions based on unclear criteria when facing out-of-distribution samples.

[0042] Based on the above analysis, to address the problem of open-set identification of small-sample radiation sources, this invention proposes a method that integrates a pre-trained semantic model and the Mahalanobis distance of physical fingerprints for open-set identification of small-sample radiation sources. The former ensures discrimination accuracy within a known class space, while the latter utilizes the compact distribution of known classes in the physical feature space to effectively define unknown samples in the open space. This multi-dimensional, cross-level fusion strategy lays the theoretical foundation for constructing a robust and reliable open-set identification system.

[0043] The core technology of this invention consists of two parts: First, for the problem of small sample size and feature extraction, a deep semantic feature extraction network based on transfer learning is constructed. Through a progressive two-stage fine-tuning strategy, the powerful generalization ability and global context modeling ability of the pre-trained semantic model under small sample conditions are fully utilized to achieve the learning of robust and discriminative deep feature representations from limited samples. Second, for the problem of open set recognition, a dual-stream decision framework that integrates physical fingerprint and deep semantic features is constructed. The Mahalanobis distance is used to accurately measure the consistency of the distribution of the test sample and the known class in the physical feature space. A set of refined fusion rules is designed to make collaborative decisions on the classification confidence of the deep network and the distribution discrimination results of the physical fingerprint, ultimately achieving highly reliable open set recognition.

[0044] This invention proposes a method for identifying open sets of radiation sources in small samples by integrating a pre-trained semantic model with Mahalanobis distance from physical fingerprints. The flowchart is shown below. Figure 1 As shown, the system block diagram is as follows: Figure 2 As shown, current methods for identifying individual radiation sources face challenges such as difficulties in small-sample learning and risks associated with open-set identification. This paper proposes a progressive training strategy to fully leverage the generalization ability of pre-trained models to address the small-sample challenge, and constructs a dual-channel feature processing framework. Simultaneously, it utilizes deep semantic features and physical fingerprint features to achieve accurate identification. Furthermore, based on the idea of ​​multi-source information fusion, and through collaborative decision-making using deep learning confidence and Mahalanobis distance distribution criteria, it achieves accurate classification of known categories and effective rejection of unknown categories, significantly improving the reliability of open-set identification.

[0045] Step 1: Obtain the complex baseband IQ signals of different radiation source individuals, and construct the training set, validation set, and test set for closed-set identification, and the test set for open-set identification.

[0046] The complex baseband IQ signal is represented as:

[0047] (1)

[0048] in Indicates signal length. Indicates components in the same direction. Represents orthogonal components It represents the imaginary unit.

[0049] Step 2: Construct a deep fusion network and train it using a progressive "freeze-fine-tuning" strategy.

[0050] The deep fusion network architecture mainly consists of four modules: a CNN front-end processing module, a projection module, a BERT module, and a classification head module. The CNN front-end processing module extracts local spatiotemporal features from the IQ signal; this part consists of multiple layers of one-dimensional convolutions, batch normalization, ReLU activation, and pooling layers. The projection module is used for data dimension matching. The BERT module is a semantic pre-trained model based on the Transformer architecture. The BERT module leverages the powerful sequence modeling capabilities of the pre-trained BERT model to further enhance the contextual representation of features. The classification head module consists of an adaptive pooling layer and multiple fully connected layers, outputting a confidence score for each known class.

[0051] The network input is complex baseband IQ dual-channel data, represented as:

[0052] (2)

[0053] The network structure of the CNN front-end processing module is shown in Table 1.

[0054] Table 1. CNN Front-End Processing Networks

[0055]

[0056] Where batch_size is the training batch size and seq_len is the sequence length, batch_size can be set to 32, but those skilled in the art can adjust this value according to actual needs. In addition, the deep fusion network adopts a one-dimensional adaptive average pooling layer (AdaptiveAvgPool1d(16)) and fixes the output to 16 feature points, so the input length is robust.

[0057] Then, the output sequence is transposed for further processing, that is, the size [batch_size, 256, 16] is transposed to [batch_size, 16, 256], and then input into the projection module.

[0058] The network structure of the projection module is shown in Table 2.

[0059] Table 2 Projection Module Network Table

[0060]

[0061] Where hidden_size is the dimension of the hidden layer, which is fixed at 768.

[0062] The BERT module uses the open-source model bert-base-cased pre-trained weights and configuration parameters, with 12 layers, 12 self-attention heads, and an attention mask dimension of [batch_size, 16]. The input and output dimensions are both [batch_size, 16, hidden_size]. However, those skilled in the art can adjust the Transformer architecture semantic pre-trained model used according to actual needs.

[0063] Then, the output sequence is transposed for dimension transformation and further processing is performed, resulting in a sequence of [batch_size, hidden_size, 16].

[0064] The size is transposed to [batch_size, hidden_size, 16] and then input into the classification head module.

[0065] The network structure of the classification head module is shown in Table 3.

[0066] Table 3 Classification Header Module Network Table

[0067]

[0068] Where num_classes represents the total number of classes of the radiation source individuals.

[0069] The gradual strategy of "freeze-fine-tuning" for deep fusion networks includes two phases:

[0070] Phase 1: Freeze all values ​​of the BERT modules and train only the CNN front-end processing module, projection module, and classification head module to enable the model to quickly learn the basic features of the signal. In this phase, the CNN learning rate can be set to 1e-3. When the model's recognition accuracy on the closed-set validation set exceeds a preset threshold... Training terminates when the threshold is not reached. Then proceed to stage two;

[0071] Phase Two: Unfreeze the weights of the last two layers of the BERT module and fine-tune the entire network using a layered learning rate strategy to enable the model to learn more refined features. Specifically, the BERT layers can use a low learning rate of 1e-5 to stabilize their pre-trained features, the CNN layers can use a medium learning rate of 5e-4, and the classification head can maintain a learning rate of 1e-3.

[0072] The above progressive strategy can guide the model to learn fine-grained features of radiation source signals without excessively perturbing BERT's existing knowledge. Those skilled in the art can adjust the values ​​of each learning rate according to actual needs.

[0073] The network training uses cross-entropy classification loss, expressed as:

[0074] (3)

[0075] in This represents the probability of predicting the class as the corresponding category. Indicates the actual label. Indicates the sample index. Indicates category index, Represents the total number of samples. This indicates the total number of categories.

[0076] Step 3: Extract physical layer fingerprint features from the IQ signal and construct a Mahalanobis distance model based on known class samples to set the classification threshold.

[0077] The physical layer fingerprint features include IQ imbalance, carrier frequency offset, phase noise features, amplitude distortion, inter-symbol interference features, instantaneous phase features, and instantaneous amplitude features, totaling 16 dimensions. The calculation method is as follows:

[0078] Assume the input complex baseband signal is:

[0079] (4)

[0080] in Indicates signal length. Indicates components in the same direction. This represents orthogonal components.

[0081] IQ imbalance characterizes the gain and phase mismatch defects in the I and Q modulators of the transmitter, including amplitude imbalance and phase imbalance.

[0082] The method for calculating amplitude imbalance is as follows:

[0083] (5)

[0084] in , They are respectively , Standard deviation of the sequence.

[0085] The phase imbalance is calculated as follows:

[0086] (6)

[0087] Carrier frequency offset characterizes the difference in frequency stability between the transmitter and the local oscillator, including the standard deviation of the frequency offset.

[0088] The standard deviation of frequency offset is calculated as follows:

[0089] (7)

[0090] in , . This indicates a phase unwinding operation.

[0091] Phase noise characteristics characterize the statistical properties of short-term phase instability and phase jitter of an oscillator. These include the standard deviation of the phase difference, the absolute mean of the phase difference, and the skewness of the phase difference.

[0092] The standard deviation of phase difference is calculated as follows:

[0093] (8)

[0094] The absolute mean of the phase difference is calculated as follows:

[0095] (9)

[0096] The phase difference bias is calculated as follows:

[0097] (10)

[0098] in for The mean, It represents the mathematical expectation.

[0099] Amplitude distortion characterizes the nonlinear characteristics and amplitude fluctuations of a power amplifier. It includes amplitude variation coefficient, amplitude kurtosis, and peak-to-average power ratio.

[0100] The amplitude variation coefficient is calculated as follows:

[0101] (11)

[0102] in , , They are respectively The mean and standard deviation.

[0103] The method for calculating amplitude and kurtosis is as follows:

[0104] (12)

[0105] Peak-to-average power ratio:

[0106] (13)

[0107] Inter-symbol interference (ISI) characteristics characterize the inter-symbol interference introduced by channel memory effects and system filtering characteristics. These include the ratio of the first sidelobe to the main lobe, the ratio of the fifth sidelobe to the main lobe, and the correlation length.

[0108] The method for calculating the ratio of the first side lobe to the main lobe is as follows:

[0109] (14)

[0110] The method for calculating the ratio of the fifth side lobe to the main lobe is as follows:

[0111] (15)

[0112] The relevant length calculation method is as follows:

[0113] (16)

[0114] in Let be the autocorrelation function of the signal. for The effective length of the sequence, This represents the lag parameter in the autocorrelation function.

[0115] Instantaneous phase characteristics describe the statistical distribution of the phase. These include the instantaneous phase standard deviation and the instantaneous phase bias.

[0116] The instantaneous phase standard deviation is calculated as follows:

[0117] (17)

[0118] The instantaneous phase deviation is calculated as follows:

[0119] (18)

[0120] Instantaneous amplitude characteristics describe the statistical distribution of amplitude. These include the instantaneous amplitude standard deviation and instantaneous amplitude skewness.

[0121] The instantaneous amplitude standard deviation is calculated as follows:

[0122] (19)

[0123] The instantaneous amplitude deviation is calculated as follows:

[0124] (20)

[0125] Finally, each radiation source signal sample is extracted into a 16-dimensional physical fingerprint feature vector, used for subsequent Mahalanobis distance calculation and open set identification, represented as... .

[0126] The construction of the Mahalanobis distance model includes feature standardization, calculation of class statistics, regularization of covariance matrix, calculation of Mahalanobis distance, threshold setting, and classification decision process.

[0127] Feature standardization performs standardization processing on physical layer features:

[0128] (twenty one)

[0129] (twenty two)

[0130] (twenty three)

[0131] in Represents the total number of samples. Indicates sample The One characteristic, Indicates the first The mean of each feature, Indicates the first The standard deviation of each feature It is used for subscript distinction and has no real meaning.

[0132] Categorical statistics include the calculation of the mean vector and covariance matrix. For each category... ,have One sample.

[0133] The mean vector is calculated as follows:

[0134] (twenty four)

[0135] in The dimension representing the feature vector. Used for subscript distinction, it has no real meaning, and has the following characteristics:

[0136] (25)

[0137] The covariance matrix is ​​calculated as follows:

[0138] (26)

[0139] in:

[0140] (27)

[0141] (28)

[0142] in , represents the feature index.

[0143] The covariance matrix regularization is calculated as follows:

[0144] (29)

[0145] The Mahalanobis distance formula is calculated as follows:

[0146] (30)

[0147] Expanded to:

[0148] (31)

[0149] in This represents the feature vector of the sample to be tested. The class represented by the known samples when calculating Mahalanobis distance.

[0150] The threshold setting method is as follows:

[0151] (32)

[0152] in satisfy:

[0153] (33)

[0154] Describing the degrees of freedom as The chi-square distribution, Indicates less than or equal to quantiles The probability value.

[0155] The classification decision-making method is as follows:

[0156] Find the category with the minimum Mahalanobis distance:

[0157] (34)

[0158] if:

[0159] (35)

[0160] Then it is classified as Otherwise, it will be classified as an unknown category.

[0161] Step 4: Input the radiation source IQ signal into the trained deep fusion network, and simultaneously input the physical fingerprint extracted from the radiation source IQ signal into the Mahalanobis distance model to obtain the classification confidence of the deep learning neural network and the confidence state and distance ratio of the Mahalanobis distance criterion in parallel.

[0162] The confidence score for classification in a deep learning neural network is represented as follows:

[0163] (36)

[0164] in Representing the A probability distribution with known probabilities This indicates the maximum confidence level.

[0165] The Mahalanobis distance ratio is calculated as follows:

[0166] (37)

[0167] in For category Threshold:

[0168] (38)

[0169] The state determination method is as follows:

[0170] (39)

[0171] in:

[0172] (40)

[0173] Step 5: Based on the combined scenario of deep fusion network confidence and Mahalanobis distance criterion, perform refined fusion decision-making and output the final open set recognition result.

[0174] The fusion decision is represented as:

[0175] (41)

[0176] in This represents the confidence threshold for deep learning neural networks. This represents the minimum Mahalanobis distance ratio. This represents the Mahalanobis distance to confidence threshold.

[0177] The statement is: If the confidence level of the neural network is greater than... When the Mahalanobis distance decision is for a known class, the final decision corresponds to the known class of the neural network decision; if the neural network confidence is greater than... Furthermore, the Mahalanobis distance decision is in the unknown class, requiring further evaluation. If the distance ratio is less than the threshold... If the neural network's confidence level is less than a certain threshold, the final decision will be the one corresponding to the known class. Otherwise, if the Mahalanobis distance's confidence level is less than a certain threshold, the final decision will be the one corresponding to the known class. Furthermore, if the Mahalanobis distance decision is a known class, and the Mahalanobis distance decision is trusted, the final decision will be the known class corresponding to the Mahalanobis distance decision; if the confidence of the neural network is less than Furthermore, the decision based on the Mahalanobis distance is of the unknown class, and the final decision is also of the unknown class.

[0178] The specific implementation method is as follows:

[0179] 1: Construct the dataset, specifically:

[0180] The WiSig dataset is a large-scale, publicly available WiFi signal dataset designed to enable individual radiation source identification when the channel and receiver are unknown. Transmitters conform to the 802.11a / g standard and operate with WiFi access points on WiFi channel 11 with a center frequency of 2462MHz and a bandwidth of 20MHz. Receiver sampling rates are 25Msps, and the data format is IQ data with 256 sampling points per sample. In this specific embodiment, the first 8 classes of individuals from the SingleDay portion of the dataset are selected as the known class training set, and the last 4 classes are selected as the unknown class test set.

[0181] 2: Construct and train a deep fusion network, specifically:

[0182] The BERT backbone architecture uses the weights of the Hugging Face open-source bert-base-cased network and its corresponding configuration file. Then, the CNN front-end, projection module, BERT intermediate layers, and linear classifier head are combined to form the complete network. The optimizer used is AdamW, which allows for weight decay. In stage one, the learning rate of the CNN front-end, projection module, and linear classifier head is 1e-3, and the weight decays to 1e-4. In stage two, the learning rate of the CNN front-end is 5e-4, and the weight decays to 1e-4. The learning rate of BERT is 1e-5, and the weight decays to 1e-5. The learning rate of the projection module and linear classifier head is 1e-3, and the weight decays to 1e-4. A decision threshold is set between the two stages. Set to 0.85, train for 25 epochs per stage, and set batch_size to 32.

[0183] Cross-entropy loss is used during classification fine-tuning training. The input is dual-channel IQ data. The training set has 8 classes with 50 data points per class, and the test set has 8 classes with 100 data points per class. The training and test sets do not overlap.

[0184] The feature extraction capabilities of the constructed deep fusion network are visualized using T-SNE, as shown below. Figure 3 As shown, compared to the baseline ResNet 1D network... Figure 4 The comparison shows that the deep fusion network constructed in this invention has denser features of the same type and more dispersed features of different types, thus exhibiting superior feature extraction and classification performance.

[0185] 3: Construct the Mahalanobis distance model, specifically as follows:

[0186] First, physical fingerprint features are extracted from the IQ data in the training set, and the mean vector and covariance matrix of each class in the training set are calculated. The confidence level of the chi-square distribution is set to 0.8, and the Mahalanobis distance confidence thresholds for each class are obtained.

[0187] 4: Input the test data to obtain the model confidence score, specifically:

[0188] The first eight test data classes are considered known classes, and the last four test data classes are considered unknown classes, with 100 samples in each class. The test samples do not overlap with the training and validation samples. After feature extraction, the samples are input into a deep fusion network and a Mahalanobis distance model, respectively, to obtain the depth confidence score and the Mahalanobis distance decision result and distance ratio.

[0189] 5: The decision results of the two models are fused for refined decision-making to obtain the open set recognition result, specifically:

[0190] The confidence threshold for the deep classification score of the neural network was set to 0.8, and the confidence distance ratio threshold was set to 3.0, to obtain the final fusion decision open set recognition result.

[0191] In this embodiment, openness is introduced to describe the difficulty of open set recognition under different experimental conditions, denoted as:

[0192] (42)

[0193] in This indicates the number of classes that will appear during training; This indicates the number of classes that will appear in the test. The higher the openness value, the greater the challenge to the robustness of the algorithm in open set recognition tasks.

[0194] In addition, this embodiment compares with several commonly used open set recognition methods, specifically: the SoftMax-T method based on threshold setting; the OpenMax method based on centroid distance extreme value fitting; the M-OSR method using Mahalanobis distance for open set recognition alone; the D-Score method based on the difference between the maximum classification score and the second largest classification score for model construction and threshold judgment for open set recognition; and the small sample radiation source individual open set recognition method that integrates pre-trained semantic model and physical fingerprint Mahalanobis distance, which is the method proposed in this invention.

[0195] The open-set recognition results are shown in Table 4. To ensure the experimental validity, all methods used the same deep fusion network—the BERT backbone network—with different data selected for each openness level, and the optimal threshold for comprehensive recognition was chosen. The closed-set recognition rate represents the proportion of correctly identified samples in the known classes of the test set; the open-set recognition rate represents the proportion of correctly identified samples in the unknown classes of the test set; and the comprehensive recognition rate represents the proportion of correctly identified samples out of the total samples.

[0196] Table 4 Open set recognition results

[0197]

[0198] Experimental results show that, under different openness conditions, the method of this invention achieves effective open set recognition while maintaining excellent closed set recognition capabilities, and its overall performance is consistently superior to the comparative methods. Furthermore, as... Figure 5 As shown, the present invention was compared with other algorithms with the number of training samples as the variable. It can be seen that under the condition of small sample size, the method of the present invention still achieves the best performance, has training sample robustness, and can effectively alleviate the small sample problem in radiation source identification.

[0199] The confusion matrix for open set recognition under an openness condition of 0.106 is as follows: Figure 6 As shown.

[0200] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Any simple modifications, alterations, and equivalent changes made to the above embodiments based on the inventive essence shall still fall within the protection scope of the present invention.

Claims

1. A method for identifying an open set of individuals of a small sample radiation source, characterized in that, It comprises the following steps: Step 1: Obtain the complex baseband IQ signals of different radiation source individuals, construct a training set, a validation set and a test set for closed set identification, and a test set for open set identification; Step 2: Construct a deep fusion network and use a "freeze-fine-tune" progressive strategy to train the deep fusion network; The deep fusion network comprises a CNN front-end processing module, a projection module, a BERT module and a classification head module, wherein the CNN front-end processing module is used to extract local space-time features from the IQ signal, and the CNN front-end processing module is composed of multiple one-dimensional convolution, batch normalization, ReLU activation and pooling layers; the projection module is used for data dimension matching; the BERT module is a semantic pre-training model of the Transformer architecture, and the BERT module uses the powerful sequence modeling capability of the pre-trained BERT model to further enhance the context representation of the features; the classification head module is composed of an adaptive pooling layer and multiple fully connected layers, and outputs the confidence score of each known class; Step 3: Extract physical layer fingerprint features from the IQ signal, and construct a Mahalanobis distance model based on known class samples to set a classification threshold; Step 4: Input the radiation source IQ signal into the trained deep fusion network, and input the physical fingerprint extracted from the radiation source IQ signal into the Mahalanobis distance model, and obtain the deep learning neural network classification confidence and the Mahalanobis distance criterion confidence state and distance ratio in parallel; Step 5: According to the combination of the deep fusion network confidence and the Mahalanobis distance criterion, perform a refined fusion decision, and output the final open set identification result; The fusion decision is: a) If the deep learning classification confidence is greater than a first threshold, and the Mahalanobis distance decision result is a known class, then the final decision is the known class corresponding to the deep learning classification result; b) If the deep learning classification confidence is greater than the first threshold, and the Mahalanobis distance decision result is an unknown class, then further judge the distance ratio: if the distance ratio is less than a second threshold, then the final decision is the known class corresponding to the deep learning classification result; otherwise, the final decision is an unknown class; c) If the deep learning classification confidence is less than the first threshold, and the Mahalanobis distance decision result is a known class, then the final decision is the known class corresponding to the Mahalanobis distance decision result; d) If the deep learning classification confidence is less than the first threshold, and the Mahalanobis distance decision result is an unknown class, then the final decision is an unknown class.

2. The method of claim 1, wherein, The "freeze-fine-tune" progressive strategy of step 2 comprises two stages: Phase one: freeze all the ownership values of the BERT module, only train the CNN front-end processing module, projection module, classification head module, when the recognition accuracy of the model on the closed set data validation set exceeds the preset threshold , the training is terminated; if the threshold is not reached , then enter phase two; Stage two: unfreeze the last two layers of network weights of the BERT module, and use a hierarchical learning rate strategy to fine-tune the entire network, so that the model learns more fine features.

3. The method of claim 1, wherein, The physical layer fingerprint features of step 3 include IQ imbalance, carrier frequency offset, phase noise feature, amplitude distortion, inter-symbol interference feature, instantaneous phase feature and instantaneous amplitude feature.

4. The method of claim 1, wherein, The Mahalanobis distance model construction in step 3 includes normalizing the physical layer fingerprint features of the known category samples, calculating the mean vector and the covariance matrix of each known category, and performing regularization processing on the covariance matrix, and finally setting a Mahalanobis distance threshold for each category based on the chi-square distribution.

5. The method of claim 1, wherein, The deep learning classification confidence in step 4 is the maximum probability value after the output layer of the deep fusion network is processed by the Softmax function; the Mahalanobis distance decision result is: calculating the Mahalanobis distance from the physical layer fingerprint features of the to-be-tested sample to each known category, if the minimum Mahalanobis distance is less than the threshold value of the corresponding category, it is determined as the corresponding known category, otherwise it is determined as an unknown category; the distance ratio is the ratio of the minimum Mahalanobis distance to the threshold value of the corresponding category.

Images