Small sample motion radiation source individual identification method and system based on migration network
By employing a small-sample recognition method using transfer networks, and pre-training the residual network structure with signals from stationary radiation sources, the problem of poor recognition performance caused by limited data from moving radiation sources is solved, achieving accurate recognition and stable training under small-sample conditions.
Patent Information
- Application Number
- CN202410950633.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2026-01-16
AI Technical Summary
In existing technologies, the limited data on moving radiation sources leads to poor recognition results, especially in deep learning where overfitting is prone to occur, making it difficult to accurately identify individual moving radiation sources.
A small-sample identification method using transfer networks is adopted. By constructing an individual radiation source identification model, pre-training it using stationary radiation source signals, retaining the model parameters of the feature extraction network, and training the transferred model, combined with an improved residual network structure, the influence of the Doppler effect is eliminated, and deep features are mined.
It achieves accurate identification of individual moving radiation sources under small sample conditions, improves the identification effect, speeds up model training, and enhances the stability and performance of the model.
Smart Images
Figure CN121350866A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of communication radiation source individual identification, and in particular to a small sample motion radiation source individual identification method and system based on a transfer network. BACKGROUND
[0002] The communication radiation source individual identification technology is a technology for identifying the identity of a communication radiation source by analyzing and extracting unique fingerprint features of the communication radiation source. Due to the rapid development of communication technology, the types and quantities of communication equipment have increased dramatically, and the communication radiation source individual identification technology can ensure the safe and stable operation of a communication network. In the process of signal processing, a signal needs to enter a circuit composed of various analog components. Due to the influence of factors such as materials, environment, and process during the production of each module component of the radiation source, the final transmitted signal will have errors compared with the ideal signal. Such errors always exist in the communication signal and are unique and difficult to counterfeit. Such unique and difficult-to-counterfeit “radiation source fingerprint features” can be used as an important basis for identifying the individual of the radiation source.
[0003] The traditional communication radiation source individual identification technology often analyzes the time domain, frequency domain, or time-frequency domain of a signal. Such an analysis method has a complex process, a single identification object, and limitations, and cannot achieve satisfactory results when facing numerous different types of radiation sources. Deep learning has the ability to automatically extract deep features and can efficiently mine subtle “fingerprint features” between different radiation sources, and is therefore often used for individual identification of communication radiation sources.
[0004] In a real use scenario, a radiation source may also be moving while transmitting a signal. When the radiation source and the receiving device move relative to each other, a Doppler effect occurs, that is, the number and frequency of the received signal wave peaks deviate from the actual signal, affecting the communication quality and causing an impact on the individual identification of the radiation source. Moreover, general radiation source signal collection is usually performed under static conditions. If the signal of a high-speed moving radiation source is to be collected, an unmanned aerial vehicle or an airplane may need to be used to carry the radiation source to collect data, which is costly and difficult, resulting in that it is difficult to obtain a motion radiation source data set. The limited data set may cause overfitting of a neural network of deep learning, and the final identification result is poor. SUMMARY
[0005] In view of the above analysis, the embodiments of the present application aim to provide a small sample motion radiation source individual identification method and system based on a transfer network, to solve the problem of poor identification effect caused by limited motion radiation data in the prior art.
[0006] In one aspect, the embodiments of the present application provide a small sample motion radiation source individual identification method based on a transfer network, which comprises:
[0007] The radiation source individual identification model comprises a feature extraction network and a sample classification layer.
[0008] The radiation source individual identification model is pre-trained by using static radiation source signal sample data.
[0009] For the pre-trained radiation source individual identification model, the model parameters of the feature extraction network are retained, and the model parameters of the sample classification layer are initialized, so as to obtain a migrated radiation source individual identification model.
[0010] The migrated radiation source individual identification model is trained by using moving radiation source signal sample data.
[0011] The to-be-identified moving radiation source signal is input into the trained radiation source individual identification model, and the identity of the to-be-identified moving radiation source is obtained.
[0012] Based on the further improvement of the above method, the feature extraction network comprises a residual network comprising a plurality of residual blocks connected in sequence.
[0013] Each residual block comprises:
[0014] A first convolution branch performs one convolution operation on input data through one convolution layer.
[0015] A second convolution branch performs three convolution operations on the input data through three convolution layers.
[0016] A first summation module adds the output of the first convolution branch and the output of the second convolution branch.
[0017] A direct connection channel performs an identity mapping on the output of the first summation module.
[0018] A third convolution channel performs three convolution operations on the output of the first summation module through three convolution layers.
[0019] A second summation module adds the output of the direct connection channel and the output of the third convolution channel and takes the sum as the output of the residual block.
[0020] Based on the further improvement of the above method, the convolution kernel size of the convolution layer of the first convolution branch is 1x1, the convolution kernel size of the three convolution layers of the second convolution branch is 1x1, and the convolution kernel size of the three convolution layers of the third convolution channel is 1x1, 3x3 and 1x1 in sequence.
[0021] Based on the further improvement of the above method, the residual block further comprises a normalization layer and an activation layer,
[0022] The normalization layer and the activation layer correspond to the convolution layer one by one,
[0023] The normalization layer is used for normalizing the data processed by the corresponding convolution layer, and the activation layer is used for activating the normalized data by a Relu function.
[0024] Based on the further improvement of the above method, the feature extraction network further comprises:
[0025] The preprocessing module performs convolution operation on the input data through the convolution layer.
[0026] The size of the convolution kernel of the convolution layer of the preprocessing module is 3*3.
[0027] Based on the further improvement of the above method, the radiation source individual identification model further comprises:
[0028] The flattening layer converts the output data of the feature extraction network into one-dimensional data and inputs the sample classification layer.
[0029] Based on the further improvement of the above method, the loss function of the radiation source individual identification model is:
[0030]
[0031] In the formula, y (i) is the true label of the i th sample, is the predicted label of the i th sample.
[0032] Based on the further improvement of the above method, the radiation source individual identification model is pre-trained by using the static radiation source signal sample data, comprising the following steps:
[0033] Obtain the static radiation source signal sample data, and divide it into a large sample training set and a large sample verification set according to a preset proportion;
[0034] The large sample training set is used to train the radiation source individual identification model round by round;
[0035] After the end of each round of training, the large sample verification set is used to verify the radiation source individual identification model, and the current model prediction accuracy is obtained;
[0036] If the change amplitude of the model prediction accuracy in the continuous preset rounds is less than a preset threshold, the training is stopped.
[0037] Based on the further improvement of the above method, the migrated radiation source individual identification model is trained by using the motion radiation source signal sample data, comprising the following steps:
[0038] Obtain motion radiation source signal sample data, and divide the motion radiation source signal sample data into a small sample training set, a small sample verification set and a small sample test set according to a preset proportion;
[0039] The small sample training set is used for training the migrated radiation source individual identification model round by round.
[0040] After each round of training ends, the small sample verification set is used for verifying the migrated radiation source individual identification model, and a current model prediction accuracy is obtained.
[0041] If the change amplitudes of the model prediction accuracies in continuous preset rounds are all lower than a preset threshold, the training is stopped, and a trained radiation source individual identification model is obtained.
[0042] The small sample test set is used for testing the trained radiation source individual identification model, a model prediction accuracy is obtained, and the performance of the trained radiation source individual identification model is evaluated according to the model prediction accuracy.
[0043] In another aspect, an embodiment of the present application provides a small sample motion radiation source individual identification system based on a migration network, and the system comprises:
[0044] A radiation source individual identification model comprises a feature extraction network and a sample classification layer.
[0045] A first training module is used for pre-training the radiation source individual identification model by using static radiation source signal sample data.
[0046] A migration module is used for retaining model parameters of the feature extraction network and initializing model parameters of the sample classification layer, so as to obtain a migrated radiation source individual identification model.
[0047] A second training module is used for training the migrated radiation source individual identification model by using motion radiation source signal sample data.
[0048] An identification module is used for inputting a motion radiation source signal to be identified into the trained radiation source individual identification model for identification, so as to obtain an identity of the motion radiation source to be identified.
[0049] Compared with the prior art, the present application can at least achieve one of the following beneficial effects:
[0050] 1、The present application, using the easily obtained stationary radiation source signal to pre-train the radiation source individual identification model, so that the model has certain feature extraction ability, retains the model parameters of the feature extraction network, and then uses a limited number of moving radiation source signals to train the pre-trained model and update the model parameters, thereby obtaining the finally trained model, realizing the individual identification of moving radiation source under the condition of small sample, avoiding the model from falling into over-fitting due to small sample problem, and more accurately identifying the individual of moving radiation source, improving the identification effect, and speeding up the model training process, making the model training process more stable.
[0051] 2、The present application improves the structure of the radiation source individual identification model, and constructs the radiation source individual identification model based on the efficient residual network (Resnet network) structure, eliminates the influence of Doppler effect, and excavates the deep features of the radiation source signal, which is beneficial to improve the accuracy of the individual identification of moving radiation source.
[0052] 3、The present application also improves the structure of the residual block of the residual network, simplifies the training process through twice residual addition operation, enhances the representation ability of the model, and improves the depth expandability of the network, thereby improving the performance of the radiation source individual identification model.
[0053] In the present application, the above technical solutions can be combined with each other to realize more preferred combination schemes. Other features and advantages of the present application will be described in the subsequent specification, and some advantages will become apparent from the specification, or will be understood by implementing the present application. The purpose and other advantages of the present application can be realized and obtained from the contents specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF DRAWINGS
[0054] The accompanying drawings are only for the purpose of illustrating specific embodiments and are not considered as limitations of the present application, and in the whole drawings, the same reference signs represent the same parts.
[0055] Figure 1 The flow chart of the small sample moving radiation source individual identification method based on the migration network of the embodiment of the present application;
[0056] Figure 2 The structure diagram of the radiation source individual identification model of the embodiment of the present application;
[0057] Figure 3 The principle diagram of the small sample moving radiation source individual identification method based on the migration network of the embodiment of the present application;
[0058] Figure 4This is a graph showing the accuracy changes of the training set and validation set during the training process when the model is directly trained using a dataset of moving radiation sources under small sample conditions according to an embodiment of the present invention.
[0059] Figure 5 This is a graph showing the accuracy changes of the training set and validation set during the training process, using a dataset of moving radiation sources to train the transferred model under small sample conditions in this embodiment of the invention.
[0060] Figure 6 This is the confusion matrix of the prediction results of the model on the test set, obtained by directly training the model on a dataset of moving radiation sources under small sample conditions in this embodiment of the invention.
[0061] Figure 7 The figure shown is the confusion matrix of the prediction results of the model obtained by training the transferred model on the test set under the small sample conditions of this embodiment of the invention using a dataset of moving radiation sources. Detailed Implementation
[0062] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.
[0063] One embodiment of the present invention discloses a method for identifying individual moving radiation sources in a small sample based on migration networks, such as... Figure 1 As shown. The method includes:
[0064] Step 1: Construct an individual radiation source identification model, which includes a feature extraction network and a sample classification layer;
[0065] Step 2: Pre-train the individual radiation source identification model using static radiation source signal sample data;
[0066] Step 3: For the pre-trained radiation source individual identification model, retain the model parameters of the feature extraction network and initialize the model parameters of the sample classification layer to obtain the transferred radiation source individual identification model.
[0067] Step 4: Train the individual radiation source identification model after migration using sample data of moving radiation source signals;
[0068] Step 5: Input the signal of the moving radiation source to be identified into the trained individual radiation source identification model to obtain the identity of the moving radiation source to be identified.
[0069] Compared with the prior art, in the embodiment of the present application, the radiation source individual identification model is pre-trained by using the easily obtained stationary radiation source signal, so that the model has certain feature extraction capability, the model parameters of the feature extraction network are reserved, and then the pre-trained model is trained by using a limited number of moving radiation source signals, the model parameters are updated and learned, so that the finally trained model is obtained, the moving radiation source individual identification under the condition of small sample is realized, the model is prevented from being trapped in over-fitting due to the small sample problem, the moving radiation source individual can be more accurately identified, the identification effect is improved, and the model training speed is accelerated, so that the model training process is more stable.
[0070] Step 1, constructing a radiation source individual identification model, the radiation source individual identification model comprising a feature extraction network and a sample classification layer.
[0071] The feature extraction network is used for feature extraction on input signal sample data, and the sample classification layer is used for identifying and obtaining the radiation source identity of the signal sample data according to the output of the feature extraction network.
[0072] In one embodiment, as shown in Figure 2 The feature extraction network comprises a residual network comprising a plurality of residual blocks connected in sequence.
[0073] Specifically, the residual network comprises three residual blocks connected in sequence.
[0074] In the embodiment of the present application, the structure of the radiation source individual identification model is improved, the radiation source individual identification model is constructed based on the efficient residual network structure, the influence of the Doppler effect is eliminated, the deep features of the radiation source signal are mined, and the accuracy of the moving radiation source individual identification is improved.
[0075] Each residual block comprises:
[0076] A first convolution branch performs one convolution operation on the input data through a convolution layer;
[0077] A second convolution branch performs three convolution operations on the input data through three convolution layers;
[0078] A first summation module adds the output of the first convolution branch and the output of the second convolution branch;
[0079] A direct connection channel performs an identity mapping on the output of the first summation module;
[0080] A third convolution channel performs three convolution operations on the output of the first summation module through three convolution layers;
[0081] A second summation module adds the output of the direct connection channel and the output of the third convolution channel to obtain the output of the residual block.
[0082] In the embodiment of the present application, the structure of the residual block of the residual network is also improved. Through two residual addition operations, the training process is simplified, the representation ability of the model is enhanced, and the depth scalability of the network is improved, thereby improving the performance of the radiation source individual identification model.
[0083] Specifically, the data input into the residual block is subjected to convolution operations through two parallel channels before the first addition operation. After the first addition operation, the data is again input into two parallel channels, one being a direct connection channel and the other being a convolution channel. The data passing through the direct connection channel is added again with the data passing through the convolution channel.
[0084] Specifically, the convolution kernel size of the convolution layer of the first convolution branch is 1x1, the convolution kernel size of the three convolution layers of the second convolution branch is 1x1, and the convolution kernel size of the three convolution layers of the third convolution channel is 1x1, 3x3 and 1x1 in sequence. Among them, the convolution kernel of the second convolution of the third convolution channel is set to 3x3, so as to capture more extensive spatial features.
[0085] Preferably, the residual block further comprises a normalization layer and an activation layer, wherein the normalization layer and the activation layer correspond to the convolution layer one by one. The normalization layer is used to normalize the data processed by the corresponding convolution layer, and the activation layer is used to activate the normalized data through a Relu function.
[0086] Among them, the normalization layer is arranged after each convolution layer to ensure that the input data distribution of each layer of the network remains relatively stable, thereby accelerating the training process. The activation layer is arranged to activate the normalized data through a Relu function, which can increase the expression ability of the network and alleviate the gradient vanishing problem. In this way, the use of normalization and activation function in the residual block promotes the stability and efficiency of model training, so that the deep network model can more easily learn complex feature representation.
[0087] Specifically, as shown in Figure 2 The feature extraction network further comprises a preprocessing module for performing convolution operation on the input data through a convolution layer. Among them, the convolution kernel size of the convolution layer of the preprocessing module is 3x3.
[0088] In the data initialization stage, the 3x3 convolution kernel can extract the local features of the input data, providing rich initial feature representation for the subsequent residual network, and improving the generalization ability of the model to different data. At the same time, through the 3x3 convolution kernel, the model can better adapt to different input size data.
[0089] Specifically, as shown in Figure 2 The radiation source individual identification model further comprises a flattening layer for converting the output data of the feature extraction network into one-dimensional data and inputting the one-dimensional data into the sample classification layer.
[0090] Specifically, the loss function of the radiation source individual identification model is:
[0091]
[0092] In the formula, y (i) is the true label of the i-th sample, is the predicted label of the i-th sample.
[0093] During the training process, the parameters of the model are adjusted through an optimization algorithm, such as the gradient descent method, to minimize the value of the loss function.
[0094] Specifically, the sample classification layer is a fully connected layer.
[0095] Step 2, the pre-training of the radiation source individual identification model using the static radiation source signal sample data comprises the following steps:
[0096] Obtaining static radiation source signal sample data and dividing the static radiation source signal sample data into a large sample training set and a large sample validation set according to a preset ratio;
[0097] Training the radiation source individual identification model using the large sample training set round by round;
[0098] After the end of each round of training, testing the radiation source individual identification model using the large sample validation set to obtain the current model prediction accuracy;
[0099] If the change amplitudes of the model prediction accuracy in continuous preset rounds are all lower than a preset threshold, the training is stopped.
[0100] In the prediction training process of the embodiment, the model is trained using the sample data in the large sample training set, and after the end of each training round, the model is verified using the large sample validation set to obtain the current model prediction accuracy and evaluate the model performance. If the change amplitudes of the model prediction accuracy in continuous preset rounds are all lower than a preset threshold, it indicates that the performance of the model on the validation set does not continuously improve or starts to decline, and then an early stopping mechanism is triggered to stop the training in advance, so that overfitting on the training set can be avoided while the best performance of the model on the validation set is preserved.
[0101] When the change amplitude of the prediction accuracy is lower than 0.5% in continuous 5 rounds, the training is ended in advance.
[0102] Specifically, the calculation formula of the model prediction accuracy is:
[0103]
[0104] The calculation formula of the change range of the model prediction accuracy is:
[0105] ΔA i = |A i -A i-1 |;
[0106] In the formula, ΔA i is the change range of the model prediction accuracy in the i-th round, A i is the model prediction accuracy in the i-th round, and i≥2.
[0107] Step 3, for the pre-trained radiation source individual identification model, the model parameters of the feature extraction network are retained, and the model parameters of the sample classification layer are initialized, thereby obtaining a migrated radiation source individual identification model, as shown in Figure 3 .
[0108] The feature extraction network of the model has learned rich feature representations after being pre-trained using large-scale static radiation source signal sample data. Retaining the model parameters of the feature extraction network can directly use these learned features, thereby accelerating the training process, reducing resource consumption, and improving the performance and generalization ability of the model on new tasks.
[0109] The sample classification layer is at the last layer of the model and is used to map the features to specific output categories, i.e., the identities of the radiation source individuals. Initializing the sample classification layer allows the model to learn how to adapt the feature representations learned in the pre-training stage to the new classification task, thereby improving the performance and generalization ability of the model.
[0110] It should be noted that when the model includes a preprocessing module, the weight values of the convolution kernels of the pre-trained preprocessing module should also be retained.
[0111] Step 4, the migrated radiation source individual identification model is trained using moving radiation source signal sample data, including the following steps:
[0112] Obtain moving radiation source signal sample data and divide it into a small sample training set, a small sample validation set, and a small sample test set according to a preset proportion;
[0113] Train the migrated radiation source individual identification model using the small sample training set;
[0114] Test the migrated radiation source individual identification model using the small sample validation set to obtain the prediction accuracy of the model;
[0115] If the change range of the model prediction accuracy in the continuous preset rounds is lower than the preset threshold, the training is stopped, thereby obtaining the trained radiation source individual identification model;
[0116] The trained radiation source individual identification model is tested by using the small sample test set, the prediction accuracy of the model is obtained, and the trained radiation source individual identification model is evaluated according to the prediction accuracy.
[0117] In the training process of the embodiment of the application, the model is trained by using sample data in the small sample training set, after each training round ends, the model is verified by using the small sample test set, the current model prediction accuracy is obtained, and the model performance is evaluated. If the change range of the model prediction accuracy in the continuous preset rounds is lower than the preset threshold, it indicates that the performance of the model on the verification set continuously does not improve or starts to decline, and then the early stopping mechanism is triggered, the training is stopped in advance, thereby overfitting on the training set can be avoided, and the best performance of the model on the verification set is retained. After the final training is completed, the model is tested by using the test set, and the performance of the model is finally evaluated.
[0118] When the change range of the prediction accuracy is lower than 0.5% in the continuous 5 rounds, the training is ended in advance.
[0119] Example
[0120] In the embodiment, the radiation source signal data of actual radio stations is used as a data set for small sample motion radiation source individual identification, there are seven types of radio station start-up data, mixed noise with signal-to-noise ratios of 12 dB, 15 dB and 20 dB, and a single sample is an IQ signal of 2*10240.
[0121] S0, a data set is constructed.
[0122] The first two types in the seven types are selected as a large sample data set, 790 samples in each type are selected as a large sample training set, 100 samples are selected as a large sample verification set, and 100 samples are selected as a large sample test set. The large sample data set is a static radiation source data set, and the signal does not have a Doppler effect.
[0123] The remaining five types of radio station data are selected as a small sample data set, 300 samples in each type are selected as a small sample training set, 100 samples are selected as a small sample verification set, and 100 samples are selected as a small sample test set. The small sample data set is a motion radiation source data set.
[0124] S1, a motion radiation source individual identification model is constructed based on a deep residual network (ResNet), and a specific model structure is as shown in Figure 2 .
[0125] The radiation source individual identification model comprises a feature extraction network, a flattening layer and a sample classification layer connected in sequence. The feature extraction network is used for feature extraction of input data. The flattening layer converts the output data of the feature extraction network into one-dimensional data and inputs the one-dimensional data into the sample classification layer. The sample classification layer outputs a predicted label of a signal, i.e., the identity of a signal radiation source individual. The number of prediction categories is the same as the number of sample radiation sources.
[0126] Specifically, the feature extraction network comprises a preprocessing module and three residual blocks connected in sequence.
[0127] The preprocessing module performs a convolution operation on input data through a convolution layer. The convolution kernel size of the convolution layer of the preprocessing module is 3x3.
[0128] Each residual block comprises:
[0129] A first convolution branch performs a convolution operation on input data through a convolution layer;
[0130] A second convolution branch performs three convolution operations on the input data through three convolution layers;
[0131] A first summation module adds the output of the first convolution branch to the output of the second convolution branch;
[0132] A direct connection channel performs an identity mapping on the output of the first summation module;
[0133] A third convolution channel performs three convolution operations on the output of the first summation module through three convolution layers;
[0134] A second summation module adds the output of the direct connection channel to the output of the third convolution channel and outputs the result as the output of the residual block.
[0135] The convolution kernel size of the convolution layer of the first convolution branch is 1x1, the convolution kernel size of the three convolution layers of the second convolution branch is 1x1, and the convolution kernel size of the three convolution layers of the third convolution channel is 1x1, 3x3 and 1x1 in sequence.
[0136] A distance-based cross-entropy loss function is constructed as follows:
[0137]
[0138] In the formula, y (i) is the true label of the i-th sample, is the predicted label of the i-th sample.
[0139] S2, initialize the parameters of the radiation source individual identification model.
[0140] Before pre-training using a large sample dataset, the parameters of each module of the radiation source individual identification model are randomly initialized.
[0141] S3, pre-training the model using a large sample dataset.
[0142] The radiation source individual identification model is trained round by round using the large sample training set. After the end of each round of training, the radiation source individual identification model is verified using the large sample verification set to obtain the current model prediction accuracy. An early stopping training mechanism is set for model training, and when the prediction accuracy change amplitude is less than 0.5% in continuous 5 training rounds, the training is ended in advance.
[0143] S4, for the pre-trained radiation source individual identification model, the model parameters of the feature extraction network are retained, and the model parameters of the sample classification layer are initialized, thereby obtaining the migrated radiation source individual identification model.
[0144] S5, training the migrated radiation source individual identification model using a small sample dataset.
[0145] The migrated radiation source individual identification model is trained round by round using the small sample training set. When training, the parameters of the feature extraction network are updated on the basis of the model parameters obtained by pre-training, and the parameters of the sample classification layer for the small sample dataset are trained.
[0146] After the end of each round of training, the migrated radiation source individual identification model is verified using the small sample verification set to obtain the current model prediction accuracy. An early stopping training mechanism is set for model training, and when the prediction accuracy change amplitude is less than 0.5% in continuous 5 training rounds, the training is ended in advance.
[0147] The trained radiation source individual identification model is tested using the small sample test set to obtain the model prediction accuracy, and the performance of the trained radiation source individual identification model is evaluated according to the model prediction accuracy.
[0148] Simulation experiments are performed, and the experimental results of the small sample motion radiation source individual identification method based on the migrated network of the embodiment of the present application are compared with the experimental results of the method of directly training the model using motion radiation source data under the existing small sample condition. The simulation experiment results are shown in Table 1 as follows: Figures 4-7
[0149] Figure 4 The figure shows the accuracy rate change graph of the training set and the validation set in the training process of the model directly trained by the moving radiation source data set under the small sample condition. The training stops at about 120 rounds, and the final training set accuracy rate is 99.60%, and the validation set accuracy rate is 96.40%.
[0150] Figure 5 The figure shows the accuracy rate change graph of the training set and the validation set in the training process of the model trained by the moving radiation source data set under the small sample condition. The training stops at about 50 rounds, and the final training set accuracy rate is 99.40%, and the validation set accuracy rate is 97.00%.
[0151] Comparison Figure 4 and Figure 5 As shown in the experimental results shown in the above and the below, compared with the method of directly training by the small sample data set, the training speed of the method of training by the transfer learning method of the embodiment of the present application is faster. It is found by comparison that the method of directly training by the small sample data set has a large decline in the training set and the validation set accuracy rate in the training process; the method of training by the transfer learning method of the present application does not have this situation, and the training process is more stable, and the accuracy rate of the validation set is higher.
[0152] Figure 6 The figure shows the confusion matrix of the prediction result of the model directly trained by the moving radiation source data set under the small sample condition on the test set. Finally, the average recognition accuracy rate of the test set is 95.60%.
[0153] Figure 7 The figure shows the confusion matrix of the prediction result of the model trained by the moving radiation source data set under the small sample condition on the test set. Finally, the average recognition accuracy rate of the test set is 97.80%.
[0154] Comparison Figure 6 and Figure 7 From the final test set average accuracy rate, the accuracy rate based on the method of the present application is also slightly higher than that of the method without using the transfer learning.
[0155] On the other hand, one embodiment of the present application is also a small sample moving radiation source individual identification system based on a transfer network. The system comprises:
[0156] A radiation source individual identification model comprising a feature extraction network and a sample classification layer;
[0157] A first training module for pre-training the radiation source individual identification model by using static radiation source signal sample data;
[0158] The migration module retains the model parameters of the feature extraction network and initializes the model parameters of the sample classification layer, so as to obtain a migrated radiation source individual identification model;
[0159] The second training module trains the migrated radiation source individual identification model by using the moving radiation source signal sample data.
[0160] The identification module inputs the to-be-identified moving radiation source signal into the trained radiation source individual identification model, and obtains the identity of the to-be-identified moving radiation source.
[0161] The system of the embodiment of the present application is used for realizing the small sample moving radiation source individual identification based on the migration network of the foregoing embodiment. The system and the method of the present application have the same / corresponding technical features and can realize the same technical effects, which will not be described here.
[0162] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiment methods can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. The computer readable storage medium is a disk, an optical disk, a read-only memory or a random access memory, etc.
[0163] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.
Claims
1. A method for small sample motion radiation source individual identification based on a migration network, characterized in that, The method comprises: constructing a radiation source individual identification model, the radiation source individual identification model comprising a feature extraction network and a sample classification layer; pre-training the radiation source individual identification model using static radiation source signal sample data; for the pre-trained radiation source individual identification model, retaining model parameters of the feature extraction network and initializing model parameters of the sample classification layer, thereby obtaining a migrated radiation source individual identification model; training the migrated radiation source individual identification model using moving radiation source signal sample data; inputting a to-be-identified moving radiation source signal into the trained radiation source individual identification model to obtain an identity of the to-be-identified moving radiation source.
2. The method of claim 1, wherein, The feature extraction network comprises: a residual network comprising a plurality of residual blocks connected in sequence; each residual block comprises: a first convolution branch performing one convolution operation on input data through a convolution layer; a second convolution branch performing three convolution operations on the input data through three convolution layers; a first summation module adding outputs of the first convolution branch and the second convolution branch; a direct connection channel performing an identity mapping on an output of the first summation module; a third convolution channel performing three convolution operations on the output of the first summation module through three convolution layers; a second summation module adding outputs of the direct connection channel and the third convolution channel to obtain an output of the residual block.
3. The method of claim 2, wherein, The convolution kernel size of the convolution layer of the first convolution branch is 1x1, the convolution kernel size of the three convolution layers of the second convolution branch is 1x1, and the convolution kernel size of the three convolution layers of the third convolution channel is 1x1, 3x3 and 1x1 in sequence.
4. The method of claim 2, wherein, The residual block further comprises a normalization layer and an activation layer, wherein the normalization layer and the activation layer correspond to the convolution layer one by one, the normalization layer is configured to normalize data processed by the corresponding convolution layer, and the activation layer is configured to perform an activation operation on the normalized data through a Relu function.
5. The method according to any one of claims 1-4, characterized in that, The feature extraction network further comprises: a preprocessing module configured to perform a convolution operation on input data through a convolution layer; wherein the convolution kernel size of the convolution layer of the preprocessing module is 3x3.
6. The method according to any one of claims 1-4, characterized in that, The radiation source individual identification model further comprises: a flattening layer configured to convert output data of the feature extraction network into one-dimensional data and input the one-dimensional data into the sample classification layer.
7. The method according to any one of claims 1-4, characterized in that, The loss function of the radiation source individual identification model is: In the formula, y (i) is a true label of the i-th sample, is a predicted label of the i-th sample.
8. The method according to any one of claims 1-4, characterized in that, The pre-training of the radiation source individual identification model using static radiation source signal sample data comprises the following steps: obtaining static radiation source signal sample data and dividing the static radiation source signal sample data into a large sample training set and a large sample validation set according to a preset ratio; training the radiation source individual identification model using the large sample training set round by round; after the end of each round of training, verifying the radiation source individual identification model using the large sample validation set to obtain a current model prediction accuracy; if the change amplitudes of the model prediction accuracy in continuous preset rounds are all lower than a preset threshold, stopping the training.
9. The method according to any one of claims 1-4, characterized in that, The migrated radiation source individual identification model is trained by using the moving radiation source signal sample data, including the following steps: Obtain moving radiation source signal sample data, and divide it into a small sample training set, a small sample verification set and a small sample test set according to a preset proportion; The migrated radiation source individual identification model is trained by using the small sample training set in each round; After each round of training, the migrated radiation source individual identification model is verified by using the small sample verification set, and the current model prediction accuracy is obtained; If the change amplitude of the model prediction accuracy in the continuous preset rounds is less than a preset threshold, the training is stopped, and a trained radiation source individual identification model is obtained; The trained radiation source individual identification model is tested by using the small sample test set, the model prediction accuracy is obtained, and the performance of the trained radiation source individual identification model is evaluated according to the model prediction accuracy.
10. A small sample moving radiation source individual identification system based on a migration network, characterized in that, The system comprises: A radiation source individual identification model comprising a feature extraction network and a sample classification layer; A first training module for pre-training the radiation source individual identification model by using static radiation source signal sample data; A migration module for retaining the model parameters of the feature extraction network and initializing the model parameters of the sample classification layer, thereby obtaining a migrated radiation source individual identification model; A second training module for training the migrated radiation source individual identification model by using moving radiation source signal sample data; An identification module for inputting a to-be-identified moving radiation source signal into the trained radiation source individual identification model, and obtaining the identity of the to-be-identified moving radiation source.