A method, medium, and device for individual radiation source identification based on contrastive learning

Abstract

This invention provides a method, medium, and apparatus for individual radiation source identification based on contrastive learning. The input data is the IQ data of radiation source signals. Feature extraction is performed through a radiation source deep feature extraction network based on contrastive learning. The contrastive learning method ensures that feature vectors of the same radiation source are as close as possible in the embedding space, while feature vectors of different radiation sources are as far apart as possible. Finally, distance calculation and evaluation are performed by comparing the distance between the predicted signal feature vector and the feature vectors in the database to obtain the individual category of the radiation source. This invention combines individual radiation source identification with contrastive learning, ensuring that feature vectors of the same radiation source are as close as possible in the embedding space, while feature vectors of different radiation sources are as far apart as possible, thereby achieving clustering of radiation sources in a high-dimensional space and realizing the identification of individual radiation sources.

Description

Technical Field

[0001] This invention relates to the field of radiation source individual identification technology, and more specifically, to a radiation source individual identification method, medium, and device based on contrastive learning. Background Technology

[0002] As modern warfare evolves towards informatization, electronic warfare plays an increasingly important role. Among its key technologies, specific radiation source identification is crucial in electronic warfare, particularly in communication reconnaissance under non-cooperative communication conditions. This involves identifying the radiation source based on subtle differences in intercepted radio frequency communication signals, and understanding the enemy's equipment carrier based on the radiation source's operational status, thereby enabling targeted surveillance, jamming, and strikes. Currently, research on radiation source identification both domestically and internationally primarily focuses on two types of signals: transient and steady-state signals. Transient signals, generated when a radiation source is switched on or off due to instability, reveal the source's characteristics. Because of their short duration and susceptibility to noise interference, they are difficult to identify. Steady-state signals, radiated when a radiation source is operating stably, are easier to obtain and less affected by noise interference, making them the focus of radiation source identification research.

[0003] Traditional radiation source identification methods can only identify known radiation sources individually. They are ineffective or even impossible to identify unknown radiation sources. Even with common deep learning methods, the absence of unknown sources in the training set leads to false positives. Therefore, effectively identifying unknown radiation sources has become a significant research challenge in radiation source identification. Summary of the Invention

[0004] The present invention aims to provide a radiation source individual identification method, medium and device based on contrastive learning, in order to solve the problem that traditional radiation source individual identification can only identify known radiation sources, and the identification effect of unknown radiation source individuals is poor or even impossible. Even when using common deep learning methods, the unknown radiation sources are not included in the training set, which directly leads to the problem of false detection.

[0005] The present invention provides a method for individual radiation source identification based on contrastive learning, comprising the following steps:

[0006] Step 1: Input and save K radiation source signals from the sensor;

[0007] Step 2: First, split the radiation source signal into I-channel data and Q-channel data, and then divide it into segments of 1024.

[0008] Step 3: Normalize each group of I-path and Q-path data after segmentation;

[0009] Step 4: Use the radiation source signals with normalized I-channel and Q-channel data as training data, and perform one-hot encoding according to the numbers 1 to M;

[0010] Step 5: Combine the training data into triples ((A i A q ),(B i B q ),(C i C q )) as training units; where A i A q For the I-channel data and Q-channel data of the reference element, B i B q For the reference element and other random signals under the same transmitter, C i C q For the reference element, random other signals from different transmitters are recorded, and the transmitter information of the triplet unit is recorded;

[0011] Step 6: Perform matrix processing on the data processed in Step 5 according to the dimensions of N*3*2*1024, then concatenate the classification labels according to the receiver information corresponding to the data, and perform N*M matrix processing to obtain the dataset; where N represents the number of data, 3 is the number of training units, 2 is the I / Q data, 1024 is the length of the I / Q data, and M represents the number of bits of one-hot encoding;

[0012] Step 7: Feed the dataset into the radiation source deep feature extraction network based on contrastive learning for training. Training ends after the loss converges, and the model weight file is obtained.

[0013] Step 8: Input the normalized IQ data of the radiation source signal in the training data into the radiation source deep feature extraction network based on contrastive learning, which loads the model weight file, to obtain the feature vector of the radiation source signal in the training data. Store the feature vector in the database according to the radiation source category.

[0014] Step 9: Input the normalized IQ data of the radiation source signal to be predicted into the radiation source deep feature extraction network based on contrastive learning with the weight file loaded to obtain the feature vector of the radiation source signal to be predicted. Calculate the Euclidean distance between the feature vector and all feature vectors in the database. If the minimum distance of the matching feature vector is less than the threshold, the radiation source signal to be predicted belongs to the category of the matching feature vector. If the minimum distance of the matching feature vector is greater than or equal to the threshold, the radiation source signal to be predicted is considered a new radiation source category, and the feature vector is automatically added to the database.

[0015] Furthermore, in step 2, when dividing the data according to a length of 1024, the I-path data and Q-path data that are less than 1024 need to be discarded.

[0016] Further, the method for normalization in step 3 is as follows:

[0017]

[0018]

[0019] where represents the normalized I-channel data and Q-channel data; D(I,Q) represents the I-channel data and Q-channel data before normalization; min() represents taking the minimum value, and max() represents taking the maximum value.

[0020] Further, in step 7, the data set needs to be divided into a training set, a validation set, and a test set according to a ratio.

[0021] Further, in step 7, the loss function during training is the multi-class cross-entropy loss, which is expressed as:

[0022]

[0023]

[0024] In the formula, are two samples, D W is and in the Euclidean distance in the latent variable space. Y = 0 represents similar samples, and at this time, it is required that D W is as close to 0 as possible; Y = 1 represents distant samples, and at this time, it is required that D W is as close to m as possible when < m, but once the distance exceeds m, it fails and will not continue to update the parameters to make the distance larger.

[0025] Further, in step 7, the radiation source depth feature extraction network based on contrast learning is designed as follows:

[0026] The I-channel and Q-channel data are processed by a Block_1*3_64 convolutional module, a 1*2 max pooling module, and a Block_1*3_128 convolutional module to extract features, resulting in two 1*512*128 feature matrices. These two 1*512*128 feature matrices are then concatenated to obtain a 1*512*256 feature matrix. Finally, the 1*512*256 feature matrix is ​​processed by a Block_1*3_512 convolutional module and a 1*2 max pooling module to obtain a 1*256*512 feature matrix. The feature matrix is ​​processed sequentially through a Block_1*3_256 convolutional module, a 1*2 max pooling module, a Block_1*3_128 convolutional module, a 1*2 max pooling module, and a Block_1*3_32 convolutional module to obtain a 1*64*32 feature matrix. After flattening the 1*64*32 feature matrix, a 512-dimensional fully connected layer is applied to obtain a 1*512 feature vector, which is then output. Finally, a 5-dimensional fully connected layer is applied to the 1*512 feature vector to output the category.

[0027] Furthermore, in step 9, the matching method is as follows:

[0028] The Euclidean distance between the feature vector of the radiation source signal to be predicted and all feature vectors in the database is calculated. If the minimum distance between the matching feature vectors is less than the threshold, the radiation source signal to be predicted belongs to the category of the matching feature vector. If the minimum distance between the matching feature vectors is greater than or equal to the threshold, the radiation source signal to be predicted is considered to be a new radiation source category, and the feature vector is automatically added to the database.

[0029] The present invention also provides a computer terminal storage medium storing computer terminal executable instructions, which are used to execute the radiation source individual identification method based on contrastive learning as described above.

[0030] The present invention also provides a computing device, comprising:

[0031] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the contrastive learning-based individual radiation source identification method described above.

[0032] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0033] 1. This invention can effectively identify not only known radiation sources but also unknown radiation sources. While general deep learning classification techniques are effective for individual radiation source identification, they also have significant limitations. When new individual radiation sources appear, they are not included in model training, leading to misidentification. Furthermore, this method cannot guarantee that the features extracted from individual radiation sources are spatially clustered. This invention combines individual radiation source identification with contrastive learning, ensuring that feature vectors of the same radiation source are as similar as possible in the embedding space, while feature vectors of different radiation sources are as far apart as possible. This achieves clustering of radiation sources in a high-dimensional space, enabling the identification of individual radiation sources.

[0034] 2. This invention features a fast and stable feature extraction speed. In traditional modulation type recognition, there are many types of signal feature extraction methods, the process is complex, and it is difficult to effectively fuse different features. The computation time also varies depending on the situation. This invention uses deep learning methods for feature fusion and extraction, and the processing time can be controlled within 1 second under various complex conditions.

[0035] 3. This invention is simple to update and maintain, highly implementable, and has a controllable iteration cycle. Traditional modulation type identification adds strategies for specific noise or interference when encountering sudden situations, but this approach has high maintenance costs, unstable iteration cycles, and the calculation code becomes increasingly cumbersome, resulting in low maintainability. In contrast, this invention only requires adding the encountered sudden signal to the dataset for training, and then replacing the model with the new one, making maintenance convenient.

[0036] 4. This invention can learn from small sample burst radiation source data. General deep learning requires a suitable proportion of dataset to achieve good results. The small sample problem is a challenge in general deep learning for individual identification. When the sample size is small, it is impossible to learn effectively from the individual. However, the radiation source individual identification method based on contrastive learning in this invention can still perform efficient learning when dealing with small sample burst radiation sources by utilizing their spatial clustering characteristics. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is an overall flowchart of the radiation source individual identification method based on contrastive learning in an embodiment of the present invention.

[0039] Figure 2 This is a flowchart of the radiation source individual identification training process in an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of the radiation source deep feature extraction network structure based on contrastive learning designed in an embodiment of the present invention.

[0041] Figure 4 This is a diagram showing the experimental classification results of the radiation source individual identification method based on contrastive learning in an embodiment of the present invention.

[0042] Figure 5 This is a diagram showing the distance matrix between feature vectors of experimental categories in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0044] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0045] Example

[0046] like Figure 1 As shown in the figure, this embodiment proposes a radiation source individual identification method based on contrastive learning, including the following steps:

[0047] Step 1: Input and save the signals from 7 radiation sources (signals emitted by radiation sources with specific modulation schemes) from the sensor; In this embodiment, the signals emitted by radiation sources with specific modulation schemes include 16 types, including 2ASK, 2FSK, 4FSK, 8FSK, AM, FM, 2PSK, 4PSK, 8PSK, OQSK, UQPSK, 8QAM, 16QAM, 32QAM, 64QAM, and 128QAM, for a total of 16 modulation schemes;

[0048] Step 2: First, split the radiation source signal into I-channel data and Q-channel data, and then divide it into segments of 1024. The I-channel data and Q-channel data that are less than 1024 need to be discarded.

[0049] Step 3: Normalize each group of I-channel and Q-channel data after segmentation; the normalization method used in this embodiment is as follows:

[0050]

[0051]

[0052] in, This represents the normalized I-path and Q-path data; D(I,Q) represents the I-path and Q-path data before normalization; min() represents the minimum value, and max() represents the maximum value.

[0053] Step 4: Use the radiation source signals with normalized I-channel and Q-channel data as training data, and perform one-hot encoding according to the numbers 1 to M;

[0054] Step 5: Combine the training data into triples ((A i A q ),(B i B q ),(C i C q )) as training units; where A i A q For the I-channel data and Q-channel data of the reference element, B i B q For the reference element and other random signals under the same transmitter, C i C q For the reference element, random other signals from different transmitters are recorded, and the transmitter information of the triplet unit is recorded;

[0055] Step 6: Perform matrix processing on the data processed in Step 5 according to the dimensions of N*3*2*1024, then concatenate the classification labels according to the receiver information corresponding to the data, and perform N*M matrix processing to obtain the dataset; where N represents the number of data, 3 is the number of training units, 2 is the I / Q data, 1024 is the length of the I / Q data, and M represents the number of bits of one-hot encoding;

[0056] Step 7: Feed the dataset into the radiation source deep feature extraction network based on contrastive learning for training. Training ends after the loss converges, and the model weight file is obtained. The dataset needs to be divided into training set, validation set and test set according to the ratio. In this embodiment, the dataset is divided in a ratio of 7:2:1, with 0.7*N data as training set, 0.2*N data as validation set and 0.1*N data as test set.

[0057] During training, each training unit (triple unit) obtains three feature vectors through a contrastive learning-based deep feature extraction network for radiation sources. The contrastive loss for each of the three feature vectors is calculated, and the multi-class cross-entropy losses A, B, and C are calculated separately. All losses are summed to obtain the overall loss for the training data. Finally, the parameters of the features within the contrastive learning-based deep feature extraction network are updated according to the backpropagation principle. Training is complete when the losses converge, and the model weight file is output. This multi-class cross-entropy loss is expressed as:

[0058]

[0059]

[0060] In the formula, There are two samples, D W yes and In the Euclidean distance of the latent variable space, Y = 0 represents similar samples. In this case, we need to determine D... W Try to get as close to 0 as possible; Y=1 represents samples that are far apart, in which case D is required to be as close as possible to 0. W The closer the value is to m, the better. However, once the distance exceeds m, the algorithm becomes invalid and will no longer update the parameters to make the distance larger and larger.

[0061] Furthermore, the network structure of the radiation source deep feature extraction network based on contrastive learning is as follows: Figure 3 As shown, the two inputs of this radiation source deep feature extraction network based on contrastive learning are the I-channel data and Q-channel data of the radiation source signal, respectively. The size of both inputs is 1*1024. The network constructs a Block benchmark module, which first performs feature extraction on the input feature matrix using n k*k convolutional kernels. The extracted result is then used for feature extraction again using n 1*1 convolutional kernels. The two output feature matrices are added together and standardized. Finally, the feature matrix is ​​output, where (Block, 3*3, 16) has a value of 3 for k and a value of 16 for n.

[0062] Specifically: The I-channel and Q-channel data are processed using a Block_1*3_64 convolutional module, a 1*2 max-pooling module, and a Block_1*3_128 convolutional module for feature extraction, resulting in two 1*512*128 feature matrices. These two 1*512*128 feature matrices are then concatenated to obtain a 1*512*256 feature matrix. Finally, the 1*512*256 feature matrix is ​​processed using a Block_1*3_512 convolutional module and a 1*2 max-pooling module to obtain a 1*256*512 feature matrix. The feature matrix is ​​processed sequentially through a Block_1*3_256 convolutional module, a 1*2 max pooling module, a Block_1*3_128 convolutional module, a 1*2 max pooling module, and a Block_1*3_32 convolutional module to obtain a 1*64*32 feature matrix. After flattening the 1*64*32 feature matrix, a 512-dimensional fully connected layer is applied to obtain a 1*512 feature vector, which is then output. Finally, a 5-dimensional fully connected layer is applied to the 1*512 feature vector to output the category.

[0063] Step 8: Input the normalized IQ data of the radiation source signal in the training data into the radiation source deep feature extraction network based on contrastive learning, which loads the model weight file, to obtain the feature vector of the radiation source signal in the training data. Store the feature vector in the database according to the radiation source category.

[0064] Step 9: Input the normalized IQ data of the radiation source signal to be predicted into the radiation source deep feature extraction network based on contrastive learning with the weight file loaded to obtain the feature vector of the radiation source signal to be predicted. Calculate the Euclidean distance between the feature vector and all feature vectors in the database. If the minimum distance of the matching feature vector is less than the threshold, the radiation source signal to be predicted belongs to the category of the matching feature vector. If the minimum distance of the matching feature vector is greater than or equal to the threshold, the radiation source signal to be predicted is considered a new radiation source category, and the feature vector is automatically added to the database.

[0065] like Figure 4 As shown, using the method of this invention, 10,000 samples from 7 transmitters were randomly classified, achieving an accuracy of over 98%. It was observed that transmitters numbered 6 and 7, which were not included in the training, still exhibited good recognition performance. Here, id from 0 to 6 represents the radiation source numbers 1 to 7. By extracting features from the test data and calculating the mean spatial feature distance of the category features, as shown... Figure 5As shown, the distance between signals of the same category is less than 10, while the distance between signals of different categories is much greater. Spatially, radiation sources exhibit smaller distances within the same category and larger distances between different categories. Surprisingly, the signals from transmitters numbered 5 and 6, which were not involved in model training, have feature vectors with significantly larger distances to other categories than those involved in training. Based on the experimental data, we can conclude that the radiation source individual identification method based on contrastive learning of this invention can effectively extract features from radiation source signals, while ensuring that signals from the same radiation source have relatively close feature distances, while signals from different radiation sources have larger feature distances. Using feature distance for signal classification has a good effect.

[0066] Furthermore, in some embodiments, a computer terminal storage medium is proposed, storing computer terminal executable instructions for executing the radiation source individual identification method based on contrastive learning as described in the preceding embodiments. Examples of computer storage media include magnetic storage media (e.g., floppy disks, hard disks, etc.), optical recording media (e.g., CD-ROMs, DVDs, etc.) or memory such as memory cards, ROMs, or RAMs. The computer storage medium can also be distributed across a network-connected computer system, for example, as an application store.

[0067] Furthermore, in some embodiments, a computing device is proposed, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the contrastive learning-based individual radiation source identification method as described in the foregoing embodiments. Examples of computing devices include PCs, tablets, smartphones, or PDAs.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying individual radiation sources based on contrastive learning, characterized in that, Includes the following steps: Step 1: Input and save K radiation source signals from the sensor; Step 2: First, split the radiation source signal into I-channel data and Q-channel data, and then divide it into segments of 1024. Step 3: Normalize each group of I-path and Q-path data after segmentation; Step 4: Use the radiation source signals with normalized I-channel and Q-channel data as training data, and perform one-hot encoding according to the numbers 1 to M; Step 5: Combine the training data into triplet pairs As training units; among them... , For the reference element's I-channel data and Q-channel data, , For the reference element and other random signals under the same transmitter, , For the reference element, random other signals from different transmitters are recorded, and the transmitter information of the triplet unit is recorded; Step 6: Perform matrix processing on the data processed in Step 5 according to the dimensions of N*3*2*1024, then concatenate the classification labels according to the receiver information corresponding to the data, and perform N*M matrix processing to obtain the dataset; where N represents the number of data, 3 is the number of training units, 2 is the I / Q data, 1024 is the length of the I / Q data, and M represents the number of bits of one-hot encoding; Step 7: Feed the dataset into the radiation source deep feature extraction network based on contrastive learning for training. Training ends after the loss converges, and the model weight file is obtained. Step 8: Input the normalized IQ data of the radiation source signal in the training data into the radiation source deep feature extraction network based on contrastive learning, which loads the model weight file, to obtain the feature vector of the radiation source signal in the training data. Store the feature vector in the database according to the radiation source category. Step 9: Input the normalized IQ data of the radiation source signal to be predicted into the radiation source deep feature extraction network based on contrastive learning with the weight file loaded to obtain the feature vector of the radiation source signal to be predicted. Then match the feature vector of the radiation source signal to be predicted with all feature vectors in the database to obtain the radiation source category. In step 7, the radiation source deep feature extraction network based on contrastive learning is designed as follows: The I-channel and Q-channel data are processed by a Block_1*3_64 convolutional module, a 1*2 max pooling module, and a Block_1*3_128 convolutional module to extract features, resulting in two 1*512*128 feature matrices. These two 1*512*128 feature matrices are then concatenated to obtain a 1*512*256 feature matrix. Finally, the 1*512*256 feature matrix is ​​processed by a Block_1*3_512 convolutional module and a 1*2 max pooling module to obtain a 1*256*512 feature matrix. The feature matrix is ​​processed sequentially through a Block_1*3_256 convolutional module, a 1*2 max pooling module, a Block_1*3_128 convolutional module, a 1*2 max pooling module, and a Block_1*3_32 convolutional module to obtain a 1*64*32 feature matrix. After flattening the 1*64*32 feature matrix, a 512-dimensional fully connected layer is applied to obtain a 1*512 feature vector, which is then output. Finally, a 5-dimensional fully connected layer is applied to the 1*512 feature vector to output the category.

2. The radiation source individual identification method based on contrastive learning according to claim 1, characterized in that, In step 2, when dividing the data according to a length of 1024, the I-path data and Q-path data that are less than 1024 need to be discarded.

3. The radiation source individual identification method based on contrastive learning according to claim 1, characterized in that, The normalization method in step 3 is as follows: in, This represents the normalized I-path and Q-path data; This represents the I-path and Q-path data before normalization; min() means taking the minimum value, and max() means taking the maximum value.

4. The radiation source individual identification method based on contrastive learning according to claim 1, characterized in that, In step 7, the dataset needs to be divided into training, validation, and test sets proportionally.

5. The radiation source individual identification method based on contrastive learning according to claim 1, characterized in that, In step 7, the loss function during training is the multi-class cross-entropy loss, expressed as: In the formula, , There are two samples. yes and Euclidean distance in the latent variable space The representatives are similar samples, in which case it is required that Try to get as close to 0 as possible; The representative is a sample that is far apart, and in this case, it is required that... Time and The closer the better, but once the distance exceeds... It becomes invalid and will no longer update parameters to make the distance keep increasing.

6. The radiation source individual identification method based on contrastive learning according to claim 1, characterized in that, In step 9, the matching method is as follows: The Euclidean distance between the feature vector of the radiation source signal to be predicted and all feature vectors in the database is calculated. If the minimum distance between the matching feature vectors is less than the threshold, the radiation source signal to be predicted belongs to the category of the matching feature vector. If the minimum distance between the matching feature vectors is greater than or equal to the threshold, the radiation source signal to be predicted is considered to be a new radiation source category, and the feature vector is automatically added to the database.

7. A computer terminal storage medium storing computer terminal executable instructions, characterized in that, The computer terminal can execute instructions for performing the radiation source individual identification method based on contrastive learning as described in any one of claims 1-6.

8. A computing device, characterized in that, include: At least one processor; The at least one processor is also connected in communication with a memory, wherein the memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the radiation source individual identification method based on contrastive learning as described in any one of claims 1-6.

Images