Radiation source identification method, device, equipment and medium

Abstract

The present invention discloses a radiation source identification method, apparatus, device, and medium. The method comprises: obtaining a radiation signal from a radiation source to be identified; preprocessing the radiation source signal using a comparison group generator to obtain comparison group data with anchor points; inputting the comparison group data with anchor points into a pre-built and trained supervised comparison representation learning model based on a convolutional long short-term memory network (CLDNN) to obtain a feature code representing the radiation source; inputting the feature code into a classifier, classifying the radiation source features using a feature distance metric, and obtaining a radiation source identification classification result. The present invention effectively solves problems such as difficulty fitting the model with limited samples, unstable performance, and unclear gradient descent. It can fully explore and utilize deep features between samples, enabling accurate identification of radiation source data with very few samples, and is characterized by high efficiency, stability, and precision.

Description

Technical Field

[0001] The present invention relates to the technical field of radiation source identification, and in particular to a radiation source identification method, device, equipment and medium. Background Art

[0002] With the rapid development of wireless communication technology and the Internet of Things, it has become particularly important to effectively identify and verify different radiation sources. However, traditional authentication technologies, such as cryptography-based methods, cannot provide sufficient security when facing malicious users with powerful computing power. In addition, deep learning and other data-driven methods require a large amount of labeled data, so in practical applications, the difficulty and cost of obtaining these datasets are relatively high.

[0003] At present, the commonly used radiation source identification method based on the cross-entropy loss network also has some limitations. The cross-entropy loss mainly focuses on the individuality of each sample. Its training process mainly minimizes the error of each sample, while ignoring the relationship between samples. This easily leads to the learned features being unable to distinguish different categories well; moreover, the cross-entropy loss is highly dependent on the sample label and is sensitive to noise labels and outliers. If there are errors in the label, it is easy to have a negative impact on the learning results.

[0004] In addition, a possible solution is to use feature learning to map the original data into features in a high-dimensional space. A representative example is the radiation source recognition method based on the triplet loss network. This method can construct positive and negative samples by utilizing the category information of the samples, so that the feature distances of similar data are close and the feature distances of heterogeneous data are far apart. However, since contrastive learning methods like this can only consider one positive example and one negative example at a time, they cannot fully explore and utilize the information between samples, resulting in low recognition accuracy. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a radiation source identification method, device, equipment and medium to solve the technical problems of the prior art identification methods such as difficulty in model fitting under limited samples, unstable performance, and unclear gradient descent.

[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0007] In a first aspect, the present invention provides a method for identifying a radiation source, the method comprising:

[0008] Acquiring a radiation signal of a radiation source to be identified;

[0009] Preprocessing the radiation source signal by a comparison group generator to obtain comparison group data with anchor points;

[0010] The comparison group data with anchor points are input into a pre-built and trained supervised comparison representation learning model based on the convolutional long short-term memory network (CLDNN) to obtain feature encoding that characterizes the radiation source.

[0011] The feature code is input into the classifier, and the radiation source features are classified by feature distance measurement to obtain a radiation source identification and classification result.

[0012] In combination with the first aspect, optionally, the constructed supervised contrastive representation learning model based on the convolutional long short-term memory network CLDNN includes:

[0013] Three deep convolutional networks, a connection layer, a long short-term memory network LSTM layer, a fully connected layer, and a modulus normalization layer are connected in sequence; each deep convolutional network includes a convolution layer and a pooling layer;

[0014] Among them, the three deep convolutional networks are respectively used to extract different deep hidden layer features of the radiation source data; the connection layer is used to connect the features extracted by the first and third deep convolutional networks to obtain the spatial features of the radiation source data and input them into the long short-term memory network LSTM layer; the long short-term memory network LSTM layer is used to extract the dependency of features in the time series to obtain short-term time features and long-term time features; the fully connected layer is used to fuse the spatial features, short-term time features, and long-term time features of the radiation source data; the modulus normalization layer is used to normalize the fused features and output the feature code representing the radiation source.

[0015] In combination with the first aspect, optionally, the training process of the supervised contrastive representation learning model based on the convolutional long short-term memory network CLDNN includes the following steps:

[0016] Acquiring radiation signals from a preset number of radiation sources;

[0017] Preprocessing the radiation source signal by a comparison group generator to obtain comparison group data with anchor points, and integrating the comparison group data into a comparison group set as a training sample set;

[0018] Training the constructed supervised contrastive representation learning model based on the convolutional long short-term memory network CLDNN based on the training sample set;

[0019] The optimization goal is to minimize the supervised contrastive learning loss function, and the model parameters are updated through backpropagation to obtain the final trained supervised contrastive representation learning model based on convolutional long short-term memory network (CLDNN).

[0020] In combination with the first aspect, optionally, the step of preprocessing the radiation source signal by a comparison group generator to obtain comparison group data with anchor points includes:

[0021] The acquired radiation signal is cleaned to remove outliers;

[0022] The cleaned radiation signals are batched and divided into multiple comparison groups according to category labels;

[0023] A signal was randomly selected as the anchor point for each comparison group;

[0024] The connection between the same type of data in different batches is established through the anchor point to obtain the comparison group data with the anchor point.

[0025] In combination with the first aspect, optionally, the supervised contrastive learning loss function L SC The expression is as follows:

[0026]

[0027] Where Logit(i) represents the supervised contrast loss of the i-th sample, and N represents the total number of samples. The calculation formula of Logit(i) is:

[0028]

[0029]

[0030]

[0031] In the formula, Logit dot (i, j) represents the dot product similarity between the i-th sample and the j-th sample; τ represents the distillation temperature; i = 1, 2, ..., N; j = 1, 2, ..., N; z i and z j Represents the low-dimensional feature vectors of the i-th sample and the j-th sample respectively; Logit e (i, j) represents the exponential function of the dot product similarity between the i-th sample and the j-th sample; Logit e Positive example of (i,j), N + Logit e The number of positive examples for (i,j).

[0032] In conjunction with the first aspect, optionally, the step of inputting the feature code into a classifier, classifying the radiation source features by a feature distance metric, and obtaining a radiation source identification and classification result includes:

[0033] According to the input feature code and sample category label, the average low-dimensional spatial feature of each sample category is calculated and used as the standard feature code. The calculation formula is as follows:

[0034]

[0035] Where, Represents the standard feature code of the t-th class sample; t=1,2,…,T; T is the total number of sample categories; t=y n ,y n is the current sample category label of the nth input radiation source signal; z n Represents the characteristic code of the input nth radiation source signal, z n =f(x n ), x n is the input n-th radiation source signal, f(·) represents the supervised contrast representation learning network based on the convolutional long short-term memory network CLDNN;

[0036] According to the standard feature coding, the feature distance measurement is performed, and the calculation formula is:

[0037]

[0038] Where, Distance (t,j) Represents feature code z j The distance from the standard feature encoding of the t-th class sample, z j =f(x j ), z j Represents the radiation source signal x to be detected j Feature encoding;

[0039] Based on formula (6), we traverse each sample category and combine it with z j The sample class label where the standard feature code with the smallest distance is located is used as the radiation source signal x to be detected j The category to which it belongs.

[0040] In combination with the first aspect, optionally, the convolution layer includes 128 convolution kernels of size 1×8; the output dimension parameter of the long short-term memory network LSTM layer is 64; there are two fully connected layers, the first fully connected layer is set with 256 nodes, and the second fully connected layer is set with 128 nodes.

[0041] In a second aspect, the present invention provides a radiation source identification device, the device comprising:

[0042] An acquisition module, used to acquire the radiation signal of the radiation source to be identified;

[0043] a preprocessing module, configured to preprocess the radiation source signal using a comparison group generator to obtain comparison group data with anchor points;

[0044] The representation learning module is used to input the comparison group data with anchor points into a pre-built and trained supervised comparison representation learning model based on the convolutional long short-term memory network (CLDNN) to obtain feature codes that represent the radiation source;

[0045] The classification and identification module is used to input the feature code into the classifier, classify the radiation source features by feature distance measurement, and obtain the radiation source identification and classification results.

[0046] In a third aspect, the present invention provides an electronic device, including a processor and a storage medium;

[0047] The storage medium is used to store instructions;

[0048] The processor is configured to operate according to the instructions to execute the steps of the radiation source identification method as described in any one of the first aspects.

[0049] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the radiation source identification method as described in any one of the first aspects are implemented.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The present invention obtains feature codes that characterize radiation sources by inputting comparison group data with anchor points into a pre-built and trained supervised comparison representation learning model based on a convolutional long short-term memory network (CLDNN). By establishing connections between similar data in different batches through anchor points, the present invention can better utilize the relationships between samples compared to traditional radiation source identification methods, effectively solving the problems of lack of connections between samples and insufficient mining of deep features of samples under limited samples, thereby enabling accurate identification of radiation source data with very few samples.

[0052] Compared with other loss functions in representation learning, the supervised contrast loss used in the method of the present invention realizes the comparative construction of multiple positive and negative examples, solves technical problems such as difficulty in fitting the model under limited samples, unstable performance, and unclear gradient descent, and effectively improves the performance of the final recognition result.

[0053] The method of the present invention does not require additional auxiliary data sets, which makes up for the problem that existing radiation source identification methods are overly dependent on large amounts of labeled data, expands the deployment scope of radiation source identification methods, reduces the preliminary preparations for deployment, deeply mines the deep features of samples to a certain extent, increases the reliability of the network, and classifies radiation source features by combining feature distance measurement, further improving the accuracy of identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A schematic flow chart of a radiation source identification method according to an embodiment of the present invention;

[0055] Figure 2 A schematic diagram of the structure of a supervised contrastive representation learning model based on a convolutional long short-term memory network (CLDNN) provided in an embodiment of the present invention;

[0056] Figure 3 A schematic diagram of generating comparison group data with anchor points using a comparison group generator provided by an embodiment of the present invention;

[0057] Figure 4 A schematic diagram of feature distribution of a comparison group without anchor points and a comparison group with anchor points provided by an embodiment of the present invention;

[0058] Figure 5 A block diagram illustrating the implementation of data preprocessing and model fitting provided by an embodiment of the present invention;

[0059] Figure 6 Monte Carlo simulation results between the method of the present invention and the traditional supervised contrastive learning method under the condition of 5% sample size of the self-collected data set provided by the embodiment of the present invention;

[0060] Figure 7 Monte Carlo simulation results between the method of the present invention and the traditional supervised contrastive learning method under the condition of 10% sample size of the self-collected data set provided by the embodiment of the present invention;

[0061] Figure 8 This is a block diagram of the structural principles of the radiation source identification device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application. Unless there is a conflict, the embodiments of the present application and the technical features in the embodiments can be combined with each other.

[0063] The term "and / or" in this document simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, or B exists alone. Additionally, the character " / " in this document generally indicates an "or" relationship between the related objects.

[0064] Example 1:

[0065] like Figure 1 As shown, the embodiment of the present invention introduces a radiation source identification method, which specifically includes the following steps:

[0066] Step 1: Obtain the radiation signal of the radiation source to be identified;

[0067] Step 2: Preprocessing the radiation source signal by a comparison group generator to obtain comparison group data with anchor points;

[0068] Step 3: Input the comparison group data with anchor points into a pre-built and trained supervised contrast representation learning model based on the convolutional long short-term memory network (CLDNN) to obtain the feature encoding that characterizes the radiation source;

[0069] Step 4: Input the feature code into the classifier, classify the radiation source features by feature distance measurement, and obtain the radiation source identification and classification results.

[0070] As an embodiment of the present invention, Figure 2 As shown in FIG, the structure of the supervised contrastive representation learning model based on the convolutional long short-term memory network CLDNN constructed in step 3 includes: three deep convolutional networks, a connection layer, a long short-term memory network LSTM layer, a fully connected layer and a modulus normalization layer connected in sequence; each deep convolutional network includes a convolution layer and a pooling layer; wherein the convolution layer includes 128 convolution kernels of size 1×8; the output dimension parameter of the long short-term memory network LSTM layer is 64; there are two fully connected layers, the first fully connected layer is set with 256 nodes, and the second fully connected layer is set with 128 nodes.

[0071] Specifically, the dimension of the input data of this model is None×n dim×2; the data is sequentially passed through three deep convolutional networks to extract different deep hidden layer features of the radiation source data; the features extracted by the first and third deep convolutional networks are connected through the connection layer to obtain the spatial features of the radiation source data and input into the long short-term memory network LSTM layer; then the long short-term memory network LSTM layer extracts the dependency of features on the time series to obtain short-term time features and long-term time features; the spatial features, short-term time features, and long-term time features of the radiation source data are then fused through the fully connected layer; finally, the modulus normalization layer normalizes the fused features and outputs the feature code representing the radiation source, whose feature code vector z n The complex properties of the radiation source are accurately characterized.

[0072] As an embodiment of the present invention, the step of preprocessing the radiation source signal by the comparison group generator in step 2 to obtain comparison group data with anchor points includes:

[0073] Step 2.1: Clean the acquired radiation signal to remove outliers.

[0074] Step 2.2: The cleaned radiation signals are batched and divided into multiple comparison groups according to the category labels;

[0075] Step 2.3: Randomly select a signal as an anchor point for each comparison group;

[0076] Among them, a corresponding number of sample data are randomly selected from each category and sorted in the order of category. At this time, the mth comparison group can be expressed as D m ={x1,x2,...x N}={X1;X2;…;X T}; where m = 1, 2, ..., M; M is the total number of comparison groups; T is the total number of sample categories, x n is the nth radiation source signal, n=1,2,…,N; X t is a set belonging to the tth class, where t=1,2,…,T; each class of sets contains k samples. Taking the first class of sets as an example, the present invention will The last dimension in is replaced by its anchor point, is the anchor point in the first category of data; then the sum of the first category of data in the T-th comparison group can be expressed as Anchor points in the T-th category data.

[0077] Step 2.4: Establish connections between similar data in different batches through the anchor points to obtain comparison group data with anchor points.

[0078] Furthermore, after all categories of samples are processed, all the comparison group data with anchor points are formed into a complete comparison group set, which is expressed as: D min ={D1,D2,…,D M}, and use it as a training sample set.

[0079] It should be noted that in the embodiment of the present invention, the radiation source signal is converted to form a data structure called a comparison group. The construction of the comparison group is to extract the correlation between the data while ensuring the independence of each data. Each comparison group is composed of signals from multiple different radiation source device categories, where one signal in each radiation source device category is randomly selected as an anchor point. The selection of the anchor point provides a basis for ensuring the connection between the data of different comparison groups. Figure 3 and Figure 4 As shown in the figure, there is a certain connection between the anchor point signals in each comparison group and the anchor point signals of other comparison groups. It can be seen that if the anchor points are not added, the projection of similar samples between batch group A and batch group B in the feature space cannot be restricted, which will lead to a decrease in the output feature performance; and this connection is conducive to establishing meaningful comparisons between multiple comparison groups, so that the network can learn the shared characteristics and differences between different categories of radiation source equipment.

[0080] Reference Figure 5 As shown, in this embodiment of the present invention, the training process of the supervised contrastive representation learning model based on the convolutional long short-term memory network CLDNN provided in step 3 includes the following steps:

[0081] Step 3.1: Acquire radiation signals from a preset number of radiation sources;

[0082] Step 3.2: Preprocessing the radiation source signal by a comparison group generator to obtain comparison group data with anchor points, and integrating the comparison group data into a comparison group set as a training sample set;

[0083] Step 3.3: Training the constructed supervised contrastive representation learning model based on the convolutional long short-term memory network CLDNN based on the training sample set;

[0084] Step 3.4: Take the minimization of the supervised contrastive learning loss function as the optimization goal, and update the model parameters through backpropagation to obtain the final trained supervised contrastive representation learning model based on the convolutional long short-term memory network CLDNN.

[0085] Furthermore, the supervised contrastive learning loss function L provided in step 3.4 is SC The expression is as follows:

[0086]

[0087] Where Logit(i) represents the supervised contrast loss of the i-th sample, and N represents the total number of samples. The calculation formula of Logit(i) is:

[0088]

[0089]

[0090]

[0091] In the formula, Logit dot (i, j) represents the dot product similarity between the i-th sample and the j-th sample; τ represents the distillation temperature; i = 1, 2, ..., N; j = 1, 2, ..., N; z i and z j Represents the low-dimensional feature vectors of the i-th sample and the j-th sample respectively; Logit e (i, j) represents the exponential function of the dot product similarity between the i-th sample and the j-th sample; Logit e Positive example of (i,j), N + Logit e The number of positive examples for (i,j).

[0092] As an embodiment of the present invention, in step 4, the step of inputting the feature code into the classifier and classifying the radiation source features by means of feature distance measurement to obtain the radiation source identification and classification result includes:

[0093] Step 4.1: Based on the input feature code and sample category label, calculate the average low-dimensional spatial feature of each sample category and use it as the standard feature code. The calculation formula is as follows:

[0094]

[0095] Where, Represents the standard feature code of the t-th class sample; t=1,2,…,T; T is the total number of sample categories; t=y n ,y n is the current sample category label of the nth input radiation source signal; z n Represents the characteristic code of the input nth radiation source signal, z n =f(x n ), x n is the input n-th radiation source signal, f(·) represents the supervised contrast representation learning network based on the convolutional long short-term memory network CLDNN;

[0096] Step 4.2: Perform feature distance measurement based on the standard feature coding, and the calculation formula is:

[0097]

[0098] Where, Distance (t,j) Represents feature code z j The distance from the standard feature encoding of the t-th class sample, z j =f(x j ), z j Represents the radiation source signal x to be detected j Feature encoding;

[0099] Step 4.3: Based on formula (6), traverse each sample category and compare it with z j The sample class label where the standard feature code with the smallest distance is located is used as the radiation source signal x to be detected j The category to which it belongs.

[0100] It is further explained that the radiation source identification method provided by the present invention mainly considers the analysis of the electromagnetic signal of the radiation source, and uses a small amount of data to identify its unique identification features caused by subtle differences in the hardware of the radiation source transmitter; wherein, the radiation source signal is mapped into a feature vector in a high-dimensional space through a supervised comparative representation learning model based on the convolutional long short-term memory network CLDNN, and the feature vector is modulus normalized, which effectively avoids its influence on the feature similarity comparison; and, the present invention batches the input radiation source signal, divides it into multiple comparison groups according to the category label, and randomly selects a signal as an anchor point for each group to establish a connection between similar data of different batches. The system solves the problem of lack of connection between samples and insufficient mining of deep features of samples under limited samples. In addition, the present invention selects supervised contrastive learning loss to calculate the feature similarity between each feature and similar / different features, so as to evaluate the features extracted by the radiation source identification network, and updates the model parameters through reverse transmission, and then calculates the standard feature coding of each radiation source device category. The radiation source identification and classification results are obtained by combining the feature distance measurement method, which realizes the accurate identification of radiation source data with very few samples and does not require additional auxiliary data sets. It makes up for the problem that traditional radiation source identification methods are overly dependent on large amounts of data and expands the scope of application of radiation source identification methods.

[0101] To verify the effectiveness of the method of the present invention, this example tests the radiation source identification method provided by the present invention with traditional radiation source identification algorithms (cross entropy loss network and triple loss network) on self-collected datasets and public Wi-Fi datasets. A portion of the data was selected for the radiation source identification task under limited sample conditions. The specific results are as follows:

[0102] Table 1 Results of self-collected datasets

[0103]

[0104] Table 2 Experimental results of the public Wi-Fi dataset

[0105]

[0106] As shown in Tables 1 and 2, the radiation source identification method provided by the present invention achieved good results on both datasets, with improved accuracy compared to both cross-entropy loss and triplet loss networks. In the self-collected data, the accuracy of the present invention increased by 9.84% when the data ratio was 10%. In the public Wi-Fi dataset 2, the accuracy of the present invention increased by a whopping 29.24%.

[0107] In addition, Monte Carlo simulation experiments were conducted on 5% and 10% of the self-collected data sets, respectively. Figure 6 and Figure 7 As shown, the superiority of the method of the present invention over the traditional supervised contrastive learning method is verified, wherein Method 1 represents the traditional supervised contrastive learning method, Method 2 represents the method of the present invention using only the anchor point comparison group provided in the method of the present invention, Method 3 represents the method of the present invention using only the modulus normalization layer provided in the method of the present invention, and Method 4 represents the radiation source method provided by the present invention. From the simulation test results, it can be seen that the method proposed by the present invention has a significant improvement over the traditional supervised contrastive learning method. Taking 5% of the data as an example: the method of the present invention increases the median of the box plot by 9.84% by adding the anchor point comparison group, greatly improving the recognition accuracy of supervised contrastive representation learning under limited samples and improving its reliability; in addition, by adding the modulus normalization layer, the "box" of the box plot is further reduced, and its upper and lower quartile intervals are reduced from about 26% to about 5%, reducing the fluctuation of the traditional supervised contrastive representation learning model and further improving its stability.

[0108] In summary, the present invention takes into account that although the traditional supervised contrast loss method can more comprehensively utilize the information between samples through multiple positive and negative examples, it does not fully utilize the label data. At the same time, there are problems of sample separation and limited gradient return, which can easily lead to unstable recognition performance and even a sharp drop in recognition accuracy. In order to solve these problems, the radiation source identification method provided by the present invention has made important improvements to the supervised contrast loss method. Through the improved supervised contrast loss, a supervised contrast representation learning model based on the convolutional long short-term memory network CLDNN is constructed to obtain feature encoding that characterizes the radiation source. In addition, anchor points are used to construct connections between data of the same category between different comparison groups, thereby enhancing the efficiency of using label data and making up for the defects in the traditional supervised contrast loss method. In addition, the modulus of the output feature vector is limited by the modulus normalization layer, which solves the problem of limited gradient return and improves the stability of training. Therefore, the radiation source identification method provided by the present invention is efficient, stable and accurate, has important significance for the security of wireless communication technology and the Internet of Things, and is suitable for promotion and application.

[0109] Example 2:

[0110] like Figure 8 As shown, an embodiment of the present invention provides a radiation source identification device, which can be used to implement the method described in Example 1. The device includes:

[0111] An acquisition module, used to acquire the radiation signal of the radiation source to be identified;

[0112] a preprocessing module, configured to preprocess the radiation source signal using a comparison group generator to obtain comparison group data with anchor points;

[0113] The representation learning module is used to input the comparison group data with anchor points into a pre-built and trained supervised comparison representation learning model based on the convolutional long short-term memory network (CLDNN) to obtain feature codes that represent the radiation source;

[0114] The classification and identification module is used to input the feature code into the classifier, classify the radiation source features by feature distance measurement, and obtain the radiation source identification and classification results.

[0115] The radiation source identification device provided in this embodiment of the present invention and the radiation source identification method provided in Example 1 are based on the same technical concept and can produce the beneficial effects described in Example 1. For matters not fully described in this embodiment, please refer to Example 1.

[0116] Example 3:

[0117] An embodiment of the present invention provides an electronic device, including a processor and a storage medium;

[0118] The storage medium is used to store instructions;

[0119] The processor is configured to operate according to the instructions to execute the steps of any one of the methods in the first embodiment.

[0120] Example 4:

[0121] An embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the computer program implements the steps of any one of the methods in the first embodiment.

[0122] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0123] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0124] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0125] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0126] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A radiation source identification method, characterized in that: The method comprises: Acquiring a radiation signal of a radiation source to be identified; Preprocessing the radiation signal by a comparison group generator to obtain comparison group data with anchor points; The comparison group data with anchor points are input into a pre-built and trained supervised comparison representation learning model based on the convolutional long short-term memory network (CLDNN) to obtain feature encoding that characterizes the radiation source. Inputting the feature code into the classifier, classifying the radiation source features by means of feature distance measurement, and obtaining a radiation source identification and classification result; The training process of the supervised contrastive representation learning model based on the convolutional long short-term memory network CLDNN includes the following steps: Acquiring radiation signals from a preset number of radiation sources; Preprocessing the radiation signal by a comparison group generator to obtain comparison group data with anchor points, and integrating the comparison group data into a comparison group set as a training sample set; Training the constructed supervised contrastive representation learning model based on the convolutional long short-term memory network CLDNN based on the training sample set; The optimization goal is to minimize the supervised contrastive learning loss function and update the model parameters through backpropagation to obtain the final trained supervised contrastive representation learning model based on the convolutional long short-term memory network (CLDNN). The step of preprocessing the radiation signal by a comparison group generator to obtain comparison group data with anchor points includes: The acquired radiation signal is cleaned to remove outliers; The cleaned radiation signals are batched and divided into multiple comparison groups according to category labels; A signal was randomly selected as the anchor point for each comparison group; Establishing connections between similar data of different batches through the anchor points to obtain comparison group data with anchor points; The supervised contrastive learning loss function The expression is as follows: （1）， Where, Indicates the The supervised contrast loss of samples, represents the total number of samples; The calculation formula is: （2）， （3）， （4）， Where, Indicates the Samples and The dot product similarity of samples; Indicates distillation temperature; ; and Respectively represent samples and The low-dimensional feature vector of samples; Indicates the Samples and The exponential function of the dot product similarity of samples; express A positive example, express The number of positive examples; The step of inputting the feature code into a classifier and classifying the radiation source features by a feature distance metric to obtain a radiation source identification and classification result comprises: According to the input feature code and sample category label, the average low-dimensional spatial feature of each sample category is calculated and used as the standard feature code. The calculation formula is as follows: （5）， Where, Indicates the Standard feature encoding of class samples; is the total number of sample categories; , For the input The current sample category label of the radiation signal; Indicates the input The characteristic code of the radiation signal, , For the input A radiation signal, Represents a supervised contrastive representation learning network based on convolutional long short-term memory network CLDNN; According to the standard feature coding, the feature distance measurement is performed, and the calculation formula is: （6）， Where, Representation feature encoding Hedi The distance of the standard feature encoding of class samples, , Indicates the radiation signal to be detected Feature encoding; Based on formula (6), we traverse each sample category and The sample class label where the standard feature code with the smallest distance is located is used as the radiation signal to be detected The category to which it belongs.

2. The radiation source identification method according to claim 1, characterized in that: The constructed supervised contrastive representation learning model based on convolutional long short-term memory network CLDNN includes: Three deep convolutional networks, a connection layer, a long short-term memory network LSTM layer, a fully connected layer, and a modulus normalization layer are connected in sequence; each deep convolutional network includes a convolution layer and a pooling layer; Among them, the three deep convolutional networks are respectively used to extract different deep hidden layer features of the radiation source data; the connection layer is used to connect the features extracted by the first and third deep convolutional networks to obtain the spatial features of the radiation source data and input them into the long short-term memory network LSTM layer; the long short-term memory network LSTM layer is used to extract the dependency of features in the time series to obtain short-term time features and long-term time features; the fully connected layer is used to fuse the spatial features, short-term time features, and long-term time features of the radiation source data; the modulus normalization layer is used to normalize the fused features and output the feature code representing the radiation source.

3. The radiation source identification method according to claim 2, characterized in that: The convolutional layer consists of 128 layers of size The convolution kernel of the long short-term memory network LSTM layer is 64; there are two fully connected layers, the first fully connected layer is set with 256 nodes, and the second fully connected layer is set with 128 nodes.

4. A radiation source identification device, characterized in that: For implementing the radiation source identification method according to any one of claims 1 to 3, the device comprises: An acquisition module, used to acquire the radiation signal of the radiation source to be identified; a preprocessing module, configured to preprocess the radiation signal using a comparison group generator to obtain comparison group data with anchor points; The representation learning module is used to input the comparison group data with anchor points into a pre-built and trained supervised comparison representation learning model based on the convolutional long short-term memory network (CLDNN) to obtain feature codes that represent the radiation source; The classification and identification module is used to input the feature code into the classifier, classify the radiation source features by feature distance measurement, and obtain the radiation source identification and classification results.

5. An electronic device, characterized in that: including processor and storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the radiation source identification method according to any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the radiation source identification method according to any one of claims 1 to 3 are implemented.

Images