Small sample radiation source individual open set identification method

Abstract

The invention belongs to the technical field of signal processing, and discloses a small sample radiation source individual open set identification method, which comprises the following steps of: firstly, constructing a deep fusion network consisting of a CNN (Convolutional Neural Network) and a pre-trained Transform architecture semantic model, and realizing radiation source small sample transfer learning by adopting a progressive fine tuning strategy; meanwhile, fingerprint features with clear physical significance are extracted from the radiation source IQ signals, and distribution of known categories is modeled based on the mahalanobis distance. In the recognition stage, the classification confidence coefficient of the deep fusion network and the ratio of the confidence state of the mahalanobis distance model based on the physical fingerprint features to the distance are obtained in parallel, and through a designed multi-scene fusion decision rule, whether a sample belongs to a known category or is rejected to be recognized as an unknown category is cooperatively judged. According to the method, high-precision closed set classification can be realized under a small number of labeled samples, unknown radiation source individuals in an open set are effectively detected, and the practicability and reliability of radiation source individual identification in a complex electromagnetic environment are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of signal processing, and particularly relates to a small sample radiation source open set individual identification method. BACKGROUND

[0002] Radiation source individual identification and fine analysis are key tasks of cognitive electronic warfare. In the face of increasingly complex electromagnetic background and non-cooperative strong confrontation battlefield environment, the traditional radiation source individual identification method has insufficient reliability and cannot meet the actual needs; although the artificial intelligence method such as deep learning has achieved certain results, it is driven by big data and contradicts the small sample reality in actual engineering. In addition, previous radiation source research is all based on a closed set assumption, that is, the radiation source types to be identified have all appeared in the training samples. In actual classification tasks, especially in non-cooperative scenarios, sample collection becomes particularly difficult, and in most cases, the collected sample set will far fail to cover all radiation source categories, and the closed set assumption will no longer hold. Therefore, the radiation source open set individual identification technology under the condition of small samples not only needs to correctly process data, but also needs to reject unknown samples while correctly realizing known category identification, and is more meaningful to research.

[0003] The existing method has the following deficiencies:

[0004] 1. The traditional radiation source individual identification method has insufficient reliability

[0005] In the face of complex electromagnetic background and non-cooperative strong confrontation environment, the traditional method is difficult to meet the actual needs.

[0006] 2. The deep learning method has poor adaptability to small samples

[0007] Deep learning relies on a large amount of labeled data, which contradicts the "small sample" condition in actual engineering.

[0008] 3. The closed set assumption does not conform to the actual application scenario

[0009] Most existing research is based on the "closed set assumption", that is, the training samples cover all the categories to be identified; however, in a non-cooperative scenario, sample collection is difficult and it is difficult to cover all radiation source categories, and the closed set assumption does not hold.

[0010] 4. The pure data-driven method has weak discrimination ability for unknown samples

[0011] When facing categories that have not appeared before (out-of-distribution samples), there is a lack of clear and stable decision basis, and the open set identification ability is limited.

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

[0013] The existing method fails to effectively combine the semantic features extracted by deep learning and the fingerprint features with clear physical meaning, limiting the reliability and interpretability of open set identification. SUMMARY

[0014] The present application aims to provide a small sample radiation source individual open set identification method to solve the above technical problems.

[0015] To solve the above technical problems, the specific technical scheme of a small sample radiation source individual open set identification method of the present application is as follows:

[0016] A small sample radiation source individual open set identification method, 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 use a "freeze-fine-tune" progressive strategy to train the deep fusion network;

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

[0020] Step 4: Input the radiation source IQ signals into the trained deep fusion network, and input the physical fingerprint extracted from the radiation source IQ signals 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;

[0021] 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.

[0022] Further, the deep fusion network of step 2 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 signals, which are 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 Transformer architecture, which 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, which outputs the confidence score of each known class.

[0023] Further, the "freeze-fine-tune" progressive strategy of step 2 comprises two stages:

[0024] Phase one: freeze all ownership values of the BERT module, only train the CNN front-end processing module, projection module, and 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 , phase two is entered;

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

[0026] Further, the physical layer fingerprint features of 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] Further, the Mahalanobis distance model construction of step 3 includes standardizing the physical layer fingerprint features of known category samples, calculating the mean vector and 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.

[0028] Further, the deep learning classification confidence of step 4 is the maximum probability value of the output layer of the deep fusion network after Softmax function processing; the Mahalanobis distance decision result is: calculating the Mahalanobis distance of the physical layer fingerprint features of the to-be-tested sample to each known category, if the minimum 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.

[0029] Further, the fusion decision of step 5 is:

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

[0031] The small sample radiation source individual open set identification method has the following advantages:

[0032] High-precision closed set classification: through the fusion of CNN and pre-trained BERT semantic model and the progressive training strategy, high-precision differentiation between known classes can still be realized under the condition of small samples; stable and reliable open set recognition: the Mahalanobis distance is based on the signal physical layer feature, and is more sensitive to the change of unknown classes, and has high stability; intelligent decision fusion: the fusion strategy proposed can effectively combine the closed set classification ability of deep learning and the distribution out-of-distribution open set detection ability of the Mahalanobis distance, and check each other, thereby significantly reducing the probability of misjudgment of known classes and the probability of missing judgment of unknown classes; strong engineering practicability: the whole system does not need a complex model training process, and the process is clear, easy to realize and deploy. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 It is a flowchart of the present application;

[0034] Figure 2 It is a system block diagram of the present application;

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

[0036] Figure 4 It is a T-SNE visualization diagram of the ResNet 1D network;

[0037] Figure 5 It is a comparative comprehensive recognition rate curve diagram of the open set recognition algorithm;

[0038] Figure 6 It is a schematic diagram of the open set recognition confusion matrix when the openness of the present application is 0.106. DETAILED DESCRIPTION

[0039] In order to better understand the purpose, structure and function of the present application, the small sample radiation source individual open set recognition method of the present application is further described in detail below in combination with the drawings.

[0040] In recent years, the pre-training large model technology taking the Transformer architecture as the core provides a new idea for solving the small sample problem. The basic knowledge and powerful context learning ability obtained by pre-training the model on a large amount of unlabeled data can be efficiently adapted to the downstream task through transfer learning. The introduction of the model into the field of radiation source recognition can effectively overcome the dependence of traditional deep learning on large-scale labeled signals, and realize stable migration from "small sample" to "high precision". In addition, the core self-attention mechanism can perform global perception and dependency modeling on the signal sequence, thereby capturing long-distance time sequence features and deep semantic information that are difficult to learn by traditional convolutional networks, and providing more robust feature representation for individual recognition.

[0041] Meanwhile, the hardware defects contained in the radiation source signal, such as power amplifier nonlinearity, crystal oscillator deviation, etc., will show unique physical layer fingerprint characteristics (such as phase noise, frequency offset, IQ imbalance, etc.). These characteristics are strongly related to specific individuals, and have distinguishable distribution rules between different categories, which are also important features for radiation source individual identification. Such physical fingerprint characteristics have clear statistical physical meaning and distribution characteristics, and are not sensitive to model structure changes, with high stability. Therefore, in the open set identification task, the physical fingerprint provides a stable and interpretable measurement benchmark for defining "known" and "unknown" samples, effectively making up for the inherent defect of pure data-driven methods in the face of out-of-distribution samples.

[0042] In view of the above analysis, in order to solve the problem of small sample radiation source individual open set identification, the present application proposes a small sample radiation source individual open set identification method fusing pre-training semantic model and physical fingerprint Mahalanobis distance. The former ensures the discrimination accuracy in the known category space, and the latter uses the compact distribution of known classes in the physical feature space to effectively define the unknown samples in the open space. This multi-dimensional and cross-level fusion strategy lays a theoretical foundation for building a robust and reliable open set identification system.

[0043] The core technology of the present application has two parts: first, for the problem of small sample and feature extraction, a deep semantic feature extraction network based on transfer learning is constructed, which fully utilizes the strong generalization ability and global context modeling ability of the pre-training semantic model under the condition of small sample, realizes the learning of robust and discriminative deep feature representation from limited samples through the progressive two-stage fine-tuning strategy; second, for the open set identification problem, a dual-flow decision framework is constructed by fusing physical layer fingerprint and deep semantic feature, which accurately measures the distribution consistency of the to-be-measured sample and the known class in the physical feature space through Mahalanobis distance, and designs a set of refined fusion rules to make collaborative decision on the classification confidence of the deep network and the distribution discrimination result of the physical fingerprint, finally realizes the high reliability of the open set identification.

[0044] The present application proposes a small sample radiation source individual open set identification method fusing pre-training semantic model and physical fingerprint Mahalanobis distance, the flow chart is as shown in Figure 1 , and the system block diagram is as shown in Figure 2 . Based on the problems of small sample learning difficulty and open set identification risk existing in the current radiation source individual identification, it is proposed to fully tap the generalization ability of the pre-training model through the progressive training strategy to cope with the small sample challenge, and to construct a dual-channel feature processing framework, and to realize accurate identification by using deep semantic features and physical layer fingerprint features. In addition, based on the idea of multi-source information fusion, through the collaborative decision of deep learning confidence and Mahalanobis distance distribution criterion, accurate classification of known categories and effective rejection of unknown categories are realized, which significantly improves the reliability of open set identification.

[0045] Step 1: Obtain the complex baseband IQ signals of different radiation source individuals, 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] wherein represents the signal length, represents the co-directional component, represents the orthogonal component represents the imaginary unit.

[0049] Step 2: Construct a deep fusion network and use a “frozen-fine-tuning” progressive strategy to train the deep fusion network.

[0050] The deep fusion network structure mainly includes 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 is used to extract local space-time features from the IQ signal, which 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, which uses the powerful sequence modeling ability of the pre-trained BERT model to further enhance the context representation of the features; and the classification head module is composed of an adaptive pooling layer and multiple fully connected layers, which outputs the confidence score of 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 network table

[0055]

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

[0057] Then, the output sequence is dimension transposed for subsequent processing, i.e., the [batch_size, 256, 16] dimension is transposed to [batch_size, 16, 256], and then input to the projection module.

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

[0059] Table 2 Network table of projection module

[0060]

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

[0062] The BERT module adopts the pre-training weights and configuration parameters of the open source model bert-base-cased, has 12 layers, 12 self-attention heads, and the attention mask dimension is [batch_size, 16], and the input and output dimensions are both [batch_size, 16, hidden_size], but the skilled person in the art can adjust the adopted Transformer architecture semantic pre-training model according to actual needs.

[0063] Then, the output sequence is dimension transposed for subsequent processing, i.e., the [batch_size, hidden_size, 16]

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

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

[0066] Table 3 Network table of classification head module

[0067]

[0068] where num_classes is the total number of classes of the radiation source individuals.

[0069] The progressive strategy of the deep fusion network "freeze-fine-tune" includes two stages:

[0070] Stage one: freeze all weights of the BERT module, and only train the CNN front-end processing module, the projection module, and the classification head module, so that the model quickly learns the basic features of the signal. In this stage, the CNN learning rate can be set to 1e-3. When the recognition accuracy of the model on the closed set data validation set exceeds the pre-set threshold , the training is terminated; if the threshold is not reached, stage two is entered;

[0071] Phase two: unfreeze the last two layers of the BERT module, and use a hierarchical learning rate strategy to fine-tune the entire network, so that the model learns more fine-grained features. Among them, the BERT layer can use a low learning rate of 1e-5 to stabilize its pre-training features, the CNN layer can use a medium learning rate of 5e-4, and the classification head can keep a learning rate of 1e-3.

[0072] The above progressive strategy can guide the model to learn the fine-grained features of the radiation source signal without disturbing the existing knowledge of BERT. Among them, the values of each learning rate can be adjusted by those skilled in the art according to actual needs.

[0073] The network training uses cross-entropy classification loss, which is represented as:

[0074] (3)

[0075] wherein represents the probability of predicting the corresponding category, represents the actual label, represents the sample index, represents the category index, represents the total number of samples, represents the total number of category classification.

[0076] Step 3: Extract the 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 feature, amplitude distortion, inter-symbol interference feature, instantaneous phase feature, and instantaneous amplitude feature, a total of 16-dimensional fingerprint features, and the calculation method is as follows:

[0078] Suppose the input complex baseband signal is:

[0079] (4)

[0080] wherein represents the signal length, represents the same component, represents the quadrature component.

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

[0082] The calculation method of amplitude imbalance is:

[0083] (5)

[0084] wherein , respectively , standard deviation of sequence.

[0085] The phase imbalance is calculated as:

[0086] (6)

[0087] The carrier frequency offset represents the difference in frequency stability between the transmitter and the local oscillator source, including the standard deviation of the frequency offset.

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

[0089] (7)

[0090] where , . denotes the phase unwrapping operation.

[0091] The phase noise characteristics represent the statistical properties of the short-term phase instability and phase jitter of the oscillator. Including 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 the phase difference is calculated as:

[0093] (8)

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

[0095] (9)

[0096] The skewness of the phase difference is calculated as:

[0097] (10)

[0098] where is the mean of , denotes the mathematical expectation.

[0099] The amplitude distortion represents the non-linear characteristics and amplitude fluctuations of the power amplifier. Including the amplitude coefficient of variation, the amplitude kurtosis, and the peak-to-average power ratio.

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

[0101] (11)

[0102] where , , respectively The mean and standard deviation of the amplitude.

[0103] The amplitude kurtosis is calculated as:

[0104] (12)

[0105] The peak-to-average power ratio is calculated as:

[0106] (13)

[0107] The inter-symbol interference feature characterizes the channel memory effect and the inter-symbol interference introduced by the system filtering property. It includes the first side lobe to main lobe ratio, the fifth side lobe to main lobe ratio, and the correlation length.

[0108] The first side lobe to main lobe ratio is calculated as:

[0109] (14)

[0110] The fifth side lobe to main lobe ratio is calculated as:

[0111] (15)

[0112] The correlation length is calculated as:

[0113] (16)

[0114] where is the autocorrelation function of the signal, is the effective length of the sequence, denotes the lag parameter in the autocorrelation function.

[0115] The instantaneous phase feature characterizes the statistical distribution property of the phase. It includes the instantaneous phase standard deviation and the instantaneous phase skewness.

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

[0117] (17)

[0118] The instantaneous phase skewness is calculated as:

[0119] (18)

[0120] The instantaneous amplitude feature characterizes the statistical distribution property of the amplitude. It includes the instantaneous amplitude standard deviation and the instantaneous amplitude skewness.

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

[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] wherein denotes the feature index.

[0143] The covariance matrix regularization is calculated as:

[0144] (29)

[0145] The Mahalanobis distance formula is calculated as:

[0146] (30)

[0147] is expanded as:

[0148] (31)

[0149] wherein denotes the feature vector of the sample to be measured, is the class represented by the known sample when calculating the Mahalanobis distance.

[0150] The threshold setting is calculated as:

[0151] (32)

[0152] wherein satisfies:

[0153] (33)

[0154] denotes the chi-square distribution with degrees of freedom, denotes the probability value less than or equal to the quantile .

[0155] The classification decision is calculated as:

[0156] The class with the smallest Mahalanobis distance is found:

[0157] (34)

[0158] If:

[0159] (35)

[0160] then the classification is , otherwise it is classified as an unknown class.

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

[0162] The deep learning neural network classification confidence is expressed as:

[0163] (36)

[0164] wherein represents the probability distribution of the th known probability, and the maximum confidence is represented by

[0165] The Mahalanobis distance ratio is calculated as:

[0166] (37)

[0167] wherein is the threshold value of the class :

[0168] (38)

[0169] The state judgment is as follows:

[0170] (39)

[0171] wherein:

[0172] (40)

[0173] Step 5: According to the combination of the deep fusion network confidence and the Mahalanobis distance criterion, a refined fusion decision is executed, and the final open set recognition result is output.

[0174] The fusion decision is expressed as:

[0175] (41)

[0176] wherein represents the deep learning neural network confidence threshold, represents the minimum Mahalanobis distance ratio, and the Mahalanobis distance ratio confidence threshold is represented by

[0177] is expressed as: if the neural network confidence is greater than , and the Mahalanobis distance decision is the known class, the final decision is the neural network decision corresponding to the known class; if the neural network confidence is greater than , and the Mahalanobis distance decision is the unknown class, further judgment is performed. If the distance ratio is less than the threshold value , believe the neural network decision, the final decision is the known class corresponding to the neural network decision, otherwise believe the Mahalanobis distance decision, the final decision is the known class corresponding to the Mahalanobis distance decision; if the neural network confidence is less than , and the Mahalanobis distance decision is the known class, believe the Mahalanobis distance decision, the final decision is the known class corresponding to the Mahalanobis distance decision; if the neural network confidence is less than , and the Mahalanobis distance decision is the unknown class, the final decision is the unknown class.

[0178] The specific embodiment is:

[0179] 1: Construct a data set, specifically:

[0180] The WiSig data set is a large-scale public WiFi signal data set, which is committed to realizing the individual identification of radiation sources when the channel and receiver are unknown. The transmitters all comply with the 802.11a / g standard and work together with the WiFi access point on the WiFi channel 11 with the center frequency 2462MHz and the bandwidth 20MHz, the receiver sampling rate is 25Msps, and the data format is IQ data with 256 sampling points per sample. In this specific embodiment, the first 8 classes of individuals in the SingleDay part of the data set are selected as the known class training set, and the last 4 classes of individuals are selected as the unknown class test set.

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

[0182] The BERT backbone architecture adopts the Hugging Face open source bert-base-cased network weight, and adopts the network corresponding configuration file. Then the CNN front end, the projection module, the BERT middle layer and the linear classification head are combined to form a complete network. The optimizer adopts AdamW which can weight decay. In stage one, the learning rate of the CNN front end, the projection module and the linear classification head is 1e-3, and the weight decay is 1e-4; in stage two, the learning rate of the CNN front end is 5e-4, and the weight decay is 1e-4. The learning rate of BERT is 1e-5, and the weight decay is 1e-5. The learning rate of the projection module and the linear classification head is 1e-3, and the weight decay is 1e-4. The determination threshold value between the two stages is set to 0.85, each stage is trained for 25 epochs, and the batch_size is set to 32.

[0183] The cross entropy loss is used in the classification fine-tuning training, the input is the IQ data double channel data, the training set has a total of 8 classes, and each class has 50 data, the test set has a total of 8 classes, and each class has 100 data. Among them, the training set and the test set data do not coincide.

[0184] The feature extraction capability of the constructed deep fusion network is visualized by T-SNE asFigure 3 As shown, the baseline ResNet 1D network is as follows Figure 4 It can be seen that the deep fusion network constructed in the application has more dense same-class features and more dispersed different-class features, and has better feature extraction and classification performance.

[0185] 3: A Mahalanobis distance model is constructed, specifically as follows:

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

[0187] 4: The model confidence score is obtained by inputting the to-be-tested data, specifically as follows:

[0188] The first 8 classes of test set data are taken as known classes, and the last 4 classes of test set data are taken as unknown classes, each with 100 samples. The to-be-tested samples and the training samples and the validation samples do not overlap. After corresponding feature extraction, the deep fusion network and the Mahalanobis distance model are inputted to obtain the deep confidence score, the Mahalanobis distance decision result and the distance ratio.

[0189] 5: The decision results of the two models are fused for fine decision, to obtain the open set recognition result, specifically as follows:

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

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

[0192] (42)

[0193] Wherein, n represents the number of classes that will appear in training; represents the number of classes that will appear in testing. The greater the openness value, the greater the challenge of the open set recognition task to the robustness of the algorithm. In addition, the present embodiment is compared with a plurality of commonly used open set recognition methods, specifically: a SoftMax-T method based on threshold setting; an OpenMax method based on fitting of the extreme value of the centroid distance; an M-OSR method using Mahalanobis distance for open set recognition alone; a D-Score method based on the difference between the maximum classification score and the second largest classification score for model construction and threshold judgment of open set recognition; and a small sample radiation source individual open set recognition method fusing a pre-trained semantic model and a physical fingerprint Mahalanobis distance, which is the method proposed in the present application.

[0194]

[0195] ​The open set recognition results are shown in Table 4. In order to compare the rationality of the experiment, the same deep fusion network is used in each method, which is a BERT backbone network, different data is selected for each openness, and the optimal threshold for comprehensive recognition is selected. The closed set recognition rate is the proportion of samples in the known class in the test set that are recognized correctly, the open set recognition rate is the proportion of samples in the unknown class in the test set that are recognized correctly, and the comprehensive recognition rate is the proportion of samples that are recognized correctly in the total samples.

[0196] Table 4 Open set recognition results

[0197]

[0198] The experimental results show that under different openness conditions, the method of the present application maintains excellent closed set recognition ability while achieving effective open set recognition, and the comprehensive performance is always better than that of the comparison method. In addition, as shown in Table 5, the present application and other algorithms are compared with the number of training samples as a variable. It can be seen that under the condition of small samples, the method of the present application still obtains the optimal performance, has training sample robustness, and can better alleviate the small sample problem in the radiation source recognition. Figure 5

[0199] The open set recognition confusion matrix under the condition of 0.106 openness is shown in Table 6. Figure 6

[0200] The above is only a preferred embodiment of the present application, and does not limit the present application. Any simple modification, change and equivalent change of the above embodiment according to the technical essence of the present application still belongs to the protection scope of the technical solution of the present application.​​

Claims

1. A method for identifying individual open sets of radiation sources in a small sample, characterized in that, Includes the following steps: 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; Step 2: Construct a deep fusion network and train it using a progressive "freeze-fine-tuning" strategy; 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; 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. 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.

2. The method according to claim 1, characterized in that, 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 extracts local spatiotemporal features from the IQ signal and 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 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 a confidence score for each known category.

3. The method according to claim 2, characterized in that, The "freeze-fine-tune" progressive strategy described in step 2 includes two phases: 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; 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.

4. The method according to claim 1, characterized in that, 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.

5. The method according to claim 1, characterized in that, Step 3 describes the construction of the Mahalanobis distance model, which includes standardizing the physical layer fingerprint features of known class samples, calculating the mean vector and covariance matrix of each known class, regularizing the covariance matrix, and finally setting a Mahalanobis distance threshold for each class based on the chi-square distribution.

6. The method according to claim 1, characterized in that, 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.

7. The method according to claim 1, characterized in that, The fusion decision in step 5 is as follows: 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, 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 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 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 an unknown category.

Images