Open set electromagnetic radiation source individual identification method based on metric learning
Through a multi-branch architecture based on metric learning and a BNNeck dimension reduction module, the multi-scale characteristics of electromagnetic signals are extracted, and the problems of insufficient identification accuracy and poor generalization ability of electromagnetic radiation sources in the prior art are solved, achieving a more efficient identification effect.
Patent Information
- Application Number
- CN202510011252.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-03
- Publication Date
- 2025-04-18
AI Technical Summary
The prior art has problems in the identification of electromagnetic radiation source individuals with insufficient identification accuracy and poor generalization ability, especially in complex and changeable scenarios, it is difficult to effectively identify unknown radiation source individuals.
The individual recognition method of open-collection electromagnetic radiation source based on metric learning is adopted. By constructing a multi-branch architecture feature extractor, combining the BNNeck dimension reduction module and multiple loss functions, the multi-scale features of the electromagnetic signal are extracted, and the accuracy and generalization ability of the recognition are improved through the metric learning model.
It improves the accuracy and generalization ability of individual identification of electromagnetic radiation sources, and can effectively identify individuals of unknown radiation sources in complex environments to meet the needs of military and civilian tasks.
Smart Images
Figure CN119848567A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of signal processing and analysis, and is specifically an open set electromagnetic radiation source individual identification method based on metric learning. Background Art
[0002] With the rapid development of wireless communication technology and radar systems, the individual identification of electromagnetic radiation sources has become an important task in the fields of radio spectrum management, counter-terrorism, military reconnaissance, etc. The core goal of the individual identification of electromagnetic radiation sources is to accurately classify and identify the signals emitted by different types of radiation sources.
[0003] Open set recognition means that the model needs to identify unknown individuals that have not been seen during the training process. In order to cope with the situation where there are unknown radiation source individuals in actual application scenarios, open set electromagnetic radiation source individual recognition is required. At present, for the task of electromagnetic radiation source individual recognition, the combination of metric learning and feature learning is used. Applying the commonly used metric learning strategy to the design and training process of deep learning models is a hot research topic in radiation source individual recognition and open set recognition.
[0004] Metric learning improves the accuracy of open-set electromagnetic radiation source individual recognition by learning a suitable distance metric, which makes the distance between samples of the same category smaller and the distance between samples of different categories larger. Metric learning improves the accuracy of open-set electromagnetic radiation source individual recognition by learning a generalized similarity metric, which enables the model to better infer the category of new and unseen samples, making open-set electromagnetic radiation source individual recognition more advantageous in complex and difficult scenarios.
[0005] In recent years, the individual identification of electromagnetic radiation sources has attracted more and more attention from researchers, and many electromagnetic radiation source individual identification models with good performance have been proposed. At present, the existing frameworks can be roughly divided into two categories: traditional electromagnetic radiation source individual identification methods based on artificial prior knowledge and electromagnetic radiation source individual identification methods based on machine learning algorithms. Traditional electromagnetic radiation source individual identification methods usually extract information from the received electromagnetic signals through manually designed features, such as spectrum features, modulation features, time domain features, etc. Famous ones include the electromagnetic radiation source individual identification algorithm based on rectangular integral bispectrum proposed by Liang Jianghai et al. and the electromagnetic radiation source individual identification algorithm based on steady-state fingerprint features proposed by Brik V et al.
[0006] Although traditional methods for identifying individual electromagnetic radiation sources have made some progress, these assumptions and prior knowledge are all for specific scenarios, so the generalization ability is limited. That is, once the environment changes significantly, the model's ability to identify individual electromagnetic radiation sources will be significantly reduced. The machine learning-based method uses a large amount of training data to train a model for identifying individual electromagnetic radiation sources, and uses the trained model to test the test data. The machine learning-based algorithm has demonstrated strong modeling and understanding capabilities, and can extract various electromagnetic signal features.
[0007] However, these learning-based methods still have shortcomings:
[0008] 1. Traditional machine learning methods rely on classifiers (such as support vector machines, KNN, decision trees, etc.) to classify input features. These classifiers usually lack direct modeling of the relative relationship between sample categories. In open set recognition, traditional classifiers lack natural processing capabilities for unknown categories.
[0009] 2. Traditional machine learning methods perform well when the distribution of test data is consistent with that of training data, but their generalization ability is weak for data with large distribution differences, and they cannot adapt to complex environments or new signal categories. Summary of the invention
[0010] In order to address the deficiencies in the prior art, the present invention provides an open set electromagnetic radiation source individual identification method based on metric learning, in order to improve the accuracy of individual electromagnetic radiation source identification and improve the identification effect in complex and changeable scenarios, thereby providing more accurate identification results for many electromagnetic signal related tasks and meeting the needs of military and civilian tasks.
[0011] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0012] The characteristics of the open set electromagnetic radiation source individual identification method based on metric learning of the present invention are as follows:
[0013] Step 1: Obtain electromagnetic signal data sets of different radiation source individuals, preprocess them, and convert them into time-frequency diagrams;
[0014] Step 1.1: Get the ultra-short wave frequency hopping signal set of K radio stations in the same outdoor scene ;in, represents the i-th ultra-short wave frequency-hopping IQ two-way signal of the k-th radio station in the same outdoor scene; represents the number of ultra-short wave frequency-hopping IQ two-way signals collected by the kth radio station in the same outdoor scene;
[0015] At the same time, obtain The signal set of the ultra-short wave frequency hopping retrieval library of radio stations in the same outdoor scene and electromagnetic signal datasets to be identified ;in, It represents the j1th ultra-short wave frequency-hopping IQ two-way retrieval library signal of the k1th radio station in the same outdoor scene; represents the j2th IQ two-way signal to be identified of the k2th radio station in the same outdoor scene; It represents the number of ultra-short wave frequency-hopping IQ two-way retrieval library signals collected by the k1th radio station in the same outdoor scene; It indicates the number of ultra-short wave frequency-hopping IQ two-way signals to be identified collected by the k2th radio station in the same outdoor scene; ;
[0016] Step 1.2: , and Perform wavelet transform and obtain the electromagnetic signal time-frequency diagram dataset accordingly , electromagnetic signal time-frequency diagram retrieval database data set And the time-frequency diagram dataset of electromagnetic signals to be identified ;in, express Time-frequency diagram obtained by wavelet transform; express Time-frequency diagram obtained by wavelet transform; express Time-frequency diagram obtained by wavelet transform;
[0017] Step 2: Based on the multi-branch architecture, an open-set electromagnetic radiation source individual feature extractor is constructed, which includes: a common feature extraction module, a global-based branch network, a local-based branch network, a channel-based branch network, and a dimensionality reduction module;
[0018] Step 2.1: Construct a common feature extraction module, which is composed of the conv1 module and conv2 module of the OSNet backbone network, and Process and obtain The common feature map of ;
[0019] Step 2.2, construct a global branch network, which is composed of the second and third bottleneck modules, the second 2D convolution layer, the first BatchNorm2d normalization layer, the second ReLU activation function, and the global feature module in sequence;
[0020] After being processed by the second and third bottleneck modules, we get The global branch common feature map , and then processed by the second 2D convolution layer, the first BatchNorm2D normalization layer and the second ReLU activation function, it is transformed into The global branch initial feature map ;
[0021] Global feature module pair Perform 2D average pooling to obtain The global branch feature map of , and remove After the first 1 / 3 of the most activated part, Perform point multiplication, and then perform 2D maximum pooling on the result to obtain The global branch robust feature map of ;
[0022] Step 2.3, construct a local-based branch network, which is composed of the fourth and fifth bottleneck modules, the third 2D convolutional layer, the second BatchNorm2D normalization layer, the third ReLU activation function, and two branch local feature modules in sequence;
[0023] After being processed by the fourth and fifth bottleneck modules, we get The local branch common feature map , and then processed by the third 2D convolution layer, the second BatchNorm2D normalization layer and the third ReLU activation function, we get The local branch initial feature map ;
[0024] Will Input the first branch local feature module to perform 2D maximum pooling processing, and obtain The local branch global feature map ;
[0025] Will Input the second branch local feature module to perform adaptive 2D average pooling, and obtain The initial local feature map of the local branch ; then Divide it into two halves along the third dimension, and we get The first feature map of the local branch and The second feature map of the local branch ;
[0026] Step 2.4, construct a channel-based branch network, which is composed of the sixth and seventh bottleneck modules, the fourth 2D convolutional layer, the third BatchNorm2D normalization layer, the fourth ReLU activation function and the channel feature module in sequence;
[0027] After being processed by the sixth and seventh bottleneck modules, we get Channel branch common feature map , Then, after the fourth 2D convolutional layer, the third BatchNorm2D normalization layer, and the fourth ReLU activation function, the result is subjected to 2D average pooling. Initial feature map of the channel branch ;
[0028] The channel feature module will The channel dimension is divided into two feature maps, and then the channel dimensions of the two feature maps are increased by 2D convolution respectively. The first feature map of the channel branch and The second feature map of the channel branch ;
[0029] Step 2.5: Construct a dimensionality reduction module, which consists of seven BNNeck modules and , , , , , , Processing to obtain multi-branch classification scores And multi-branch comprehensive feature map ;
[0030] Step 3, using formula (1) to construct the total loss function L of the open set electromagnetic radiation source individual feature extractor;
[0031] (1)
[0032] In formula (1), represents the cross entropy loss, represents the triple loss, represents the center loss, represents the circle loss, λ1, λ2, λ3, and λ4 are cross entropy losses respectively. , triplet loss , center loss The weight and circle loss The weight of
[0033] Step 4: using an ADAM optimizer to train the open set electromagnetic radiation source individual feature extractor until the total loss function L converges, thereby obtaining an optimal open set electromagnetic radiation source individual feature extractor;
[0034] Step 5: Use the open set electromagnetic radiation source individual feature extractor to retrieve the electromagnetic signal time-frequency diagram database dataset And the time-frequency diagram dataset of electromagnetic signals to be identified Processing is performed to obtain the corresponding retrieval library signal feature set And the signal feature set to be identified ;in, They are The global branch dimensionality reduction feature map, the global branch robust dimensionality reduction feature map, the local branch global dimensionality reduction feature map, the local branch first dimensionality reduction feature map, the local branch second dimensionality reduction feature map, the channel branch first dimensionality reduction feature map, the channel branch second dimensionality reduction feature map, They are The global branch dimensionality reduction feature map, the global branch robust dimensionality reduction feature map, the local branch global dimensionality reduction feature map, the local branch first dimensionality reduction feature map, the local branch second dimensionality reduction feature map, the channel branch first dimensionality reduction feature map, and the channel branch second dimensionality reduction feature map;
[0035] Will Overlay The encoding feature map of ;
[0036] Will Overlay The encoding feature map of ;
[0037] Use formula (8) to determine and Matching results :
[0038] (8)
[0039] In formula (8), d() is the cosine distance function, It is the station number obtained by extracting features and matching reasoning from the j2th IQ two-way signal to be identified of the k2th station.
[0040] The open set electromagnetic radiation source individual identification method based on metric learning of the present invention is also characterized in that the conv1 module in step 2.1 sequentially uses the first 2D convolutional layer, the InstanceNorm2D normalization layer and the first ReLU activation function to operate on Perform feature extraction and obtain The first stage common feature map ;
[0041] The conv2 module uses the MaxPool2D layer After global maximum pooling, we get The first stage common normalized feature map , and then use the first bottleneck module to Perform feature extraction and obtain The public feature map of the second stage .
[0042] Furthermore, any bottleneck module in step 2.2, step 2.3, and step 2.4 is composed of a convolution block with a dimension of 1×1, an OS module, and a 2D average pooling layer in sequence;
[0043] Step 2.1.1: The 1×1 convolutional block in the first bottleneck module uses the fifth 2D convolutional layer, the fourth BatchNorm2D normalization layer, and the fifth ReLU activation function to operate on Perform nonlinear transformation to obtain The first stage nonlinear common feature map ;
[0044] Step 2.1.2: The OS module in the first bottleneck module uses the sixth 2D convolutional layer, the fifth BatchNorm2D normalization layer, and the sixth ReLU activation function to operate on Perform dimensionality reduction and obtain The first stage of dimensionality reduction common feature map ;
[0045] The OS module in the first bottleneck module uses the seventh 2D convolutional layer to Perform feature extraction and obtain The second stage first scale feature map ;
[0046] The OS module in the first bottleneck module uses the eighth 2D convolutional layer, the ninth 2D convolutional layer in turn. Perform feature extraction and obtain The second stage second scale feature map ;
[0047] The OS module in the first bottleneck module uses the tenth 2D convolutional layer, the eleventh 2D convolutional layer, and the twelfth 2D convolutional layer in sequence. Perform feature extraction and obtain The second stage third scale feature map ;
[0048] The OS module in the first bottleneck module uses the thirteenth 2D convolution layer, the fourteenth 2D convolution layer, the fifteenth 2D convolution layer, and the sixteenth 2D convolution layer in sequence. Perform feature extraction and obtain The second stage fourth scale feature map ;
[0049] The 2D average pooling layer in the first bottleneck module is , , , The 2D average pooling and the seventh ReLU activation function operations are performed respectively, and the results are then concatenated in the channel dimension to obtain The second stage multi-scale feature map .
[0050] Further, the step 2.5 includes:
[0051] The first BNNeck module pair After dimensionality reduction, we get The global branch dimension reduction feature map , and then use the first fully connected layer to Process and obtain The global branch dimensionality reduction classification score ;
[0052] The second BNNeck module Perform dimensionality reduction to obtain Global branch robust dimension reduction feature map , and then use the second fully connected layer to Process and obtain The global branch global dimensionality reduction classification score ;
[0053] The third BNNeck module is Perform dimensionality reduction to obtain The local branch global dimension reduction feature map , and then use the third fully connected layer to Process and obtain The local branch robust dimensionality reduction classification score ;
[0054] The fourth BNNeck module pair After dimensionality reduction, we get The first dimension reduction feature map of the local branch , and then use the fourth fully connected layer to Process and obtain The global branch first dimension reduction classification score ;
[0055] The fifth BNNeck module pair After dimensionality reduction, we get The second dimension reduction feature map of the local branch , and then use the fifth fully connected layer to Process and obtain The global branch second dimension reduction classification score ;
[0056] The sixth BNNeck module pair After dimensionality reduction, we get The first dimension reduction feature map of the channel branch , and then use the sixth fully connected layer to Process and obtain The first dimension reduction classification score of the channel branch ;
[0057] Seventh BNNeck module pair After dimensionality reduction, we get The second dimension reduction feature map of the channel branch , and then use the seventh fully connected layer to Process and obtain The second dimension reduction classification score of the channel branch ;
[0058] Will , , , , , , Overlay Multi-branch classification score ;
[0059] Will , , Overlay Multi-branch comprehensive feature map .
[0060] Furthermore, in step 3, the cross entropy loss is constructed using formula (2): :
[0061] (2)
[0062] In formula (2), x(k,i) is The symbolic function of If the true category prediction is correct, let x(k,i)=1, otherwise, let x(k,i)=0;
[0063] Use formula (3) to construct triple loss :
[0064] (3)
[0065] In formula (3), For Randomly selected from The same radio station Ultra-short wave frequency hopping IQ dual-channel signal Time-frequency diagram of The multi-branch comprehensive feature map, For Randomly selected from Not belonging to the same station Ultra-short wave frequency hopping IQ dual-channel signal Time-frequency diagram of Multi-branch comprehensive feature map; ;
[0066] Use formula (4) to construct the center loss :
[0067] (4)
[0068] In formula (4), is the mean center of the multi-branch comprehensive characteristic diagram of the time-frequency diagram of all ultra-short wave frequency-hopping IQ two-way signals belonging to the k-th radio station;
[0069] Using formula (5) to construct the circle loss :
[0070] (5)
[0071] In formula (5), is the scale factor, is the inter-class weight factor of the circular loss of the kth station, is the intra-class weight factor of the circular loss of the kth station; is the inter-class similarity of the circular loss of the kth radio station, is the intra-class similarity of the circle loss of the kth station, and:
[0072] (6)
[0073] (7)
[0074] In formula (6) and formula (7), for Middle The i-th ultra-short wave frequency-hopping IQ two-way signal of the radio station Time-frequency diagram of Multi-branch comprehensive feature map; for The i1th ultrashort wave frequency hopping IQ two-way signal of the kth station in Time-frequency diagram of Multi-branch comprehensive feature map; for The i2th ultrashort wave frequency hopping IQ two-way signal of the kth station Time-frequency diagram of Multi-branch comprehensive feature map.
[0075] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the open-collection electromagnetic radiation source individual identification method, and the processor is configured to execute the program stored in the memory.
[0076] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the open-collection electromagnetic radiation source individual identification method when the computer program is executed by a processor.
[0077] Compared with the prior art, the present invention has the following beneficial effects:
[0078] 1. The present invention constructs an open set electromagnetic radiation source individual feature extractor based on metric learning, uses the time-frequency diagram of electromagnetic signals of different radiation source individuals as input, combines the BNNeck dimension reduction module and multiple loss functions, and solves the problem that the traditional electromagnetic radiation source individual identification network simply uses classifiers for classification while ignoring the intra-class and inter-class similarities of electromagnetic signals of different radiation source individuals, thereby improving the accuracy of individual electromagnetic radiation source identification.
[0079] 2. The open set electromagnetic radiation source individual identification method based on metric learning constructed by the present invention uses a matching reasoning method to determine the identification matching result by calculating the similarity between the features extracted by the network. It is different from the traditional method of simply distinguishing between known radiation sources and unknown radiation sources, and realizes the specific classification of unknown radiation source individuals.
[0080] 3. The open set electromagnetic radiation source individual feature extractor based on metric learning constructed by the present invention adopts a multi-branch structure to extract the multi-scale features of the electromagnetic signal time-frequency diagram, obtain richer and more comprehensive time domain and frequency domain features, and enable the model to more efficiently learn important information at different scales in the signal, thereby improving the generalization ability of electromagnetic radiation source individual identification.
[0081] 4. The open set electromagnetic radiation source individual feature extractor based on metric learning constructed by the present invention introduces a bottleneck module to reduce the computational complexity; at the same time, by compressing and restoring the high-dimensional feature space, the feature extractor can capture more compact and meaningful feature representations, avoiding the interference of redundant features on the learning process, and improving the robustness and generalization ability of individual electromagnetic radiation source identification. BRIEF DESCRIPTION OF THE DRAWINGS
[0082] Figure 1 is a flow chart of the present invention;
[0083] Figure 2 It is a schematic diagram of the structure of an open set electromagnetic radiation source individual feature extractor based on metric learning of the present invention;
[0084] Figure 3 It is a schematic diagram of the structure of the bottleneck module of the open set electromagnetic radiation source individual feature extractor based on metric learning of the present invention;
[0085] Figure 4 It is a schematic diagram of the structure of the BNNeck module of the open set electromagnetic radiation source individual feature extractor based on metric learning in the present invention. DETAILED DESCRIPTION
[0086] In this embodiment, a method for individual identification of open-set electromagnetic radiation sources based on metric learning is intended to solve the problem that it is difficult to effectively distinguish known categories from unknown categories in an open environment in existing methods, as well as the limitation that traditional classification methods lack processing capabilities for unknown categories. By constructing an individual feature extractor of electromagnetic radiation sources based on metric learning, this method can learn highly adaptable feature representations and metric spaces from the features of electromagnetic radiation source signals, achieve effective rejection of unknown categories and high-precision distinction of known categories, thereby improving the performance of individual identification of radiation sources in an open-set environment and meeting the needs of high-reliability identification in complex electromagnetic environments. Specifically, Figure 1 As shown, the following steps are followed:
[0087] Step 1: Obtain electromagnetic signal data sets of different radiation source individuals, preprocess them, and convert them into time-frequency diagrams;
[0088] Step 1.1: Get the ultra-short wave frequency hopping signal set of K radio stations in the same outdoor scene ;in, represents the i-th ultra-short wave frequency-hopping IQ two-way signal of the k-th radio station in the same outdoor scene; represents the number of ultra-short wave frequency-hopping IQ two-way signals collected by the kth radio station in the same outdoor scene;
[0089] At the same time, obtain The signal set of the ultra-short wave frequency hopping retrieval library of radio stations in the same outdoor scene and electromagnetic signal datasets to be identified ;in, It represents the j1th ultra-short wave frequency-hopping IQ two-way retrieval library signal of the k1th radio station in the same outdoor scene; represents the j2th IQ two-way signal to be identified of the k2th radio station in the same outdoor scene; It represents the number of ultra-short wave frequency-hopping IQ two-way retrieval library signals collected by the k1th radio station in the same outdoor scene; It indicates the number of ultra-short wave frequency-hopping IQ two-way signals to be identified collected by the k2th radio station in the same outdoor scene; In this embodiment, , .
[0090] Step 1.2: , and Perform wavelet transform and obtain the electromagnetic signal time-frequency diagram dataset accordingly , electromagnetic signal time-frequency diagram retrieval database data set And the time-frequency diagram dataset of electromagnetic signals to be identified ;in, express Time-frequency diagram obtained by wavelet transform; express Time-frequency diagram obtained by wavelet transform; express Time-frequency diagram obtained by wavelet transform.
[0091] Step 2: Figure 2 As shown, based on the multi-branch architecture, an open set electromagnetic radiation source individual feature extractor is constructed, which includes: a common feature extraction module, a global-based branch network, a local-based branch network, a channel-based branch network and a dimensionality reduction module;
[0092] Step 2.1: Construct a common feature extraction module, which is composed of the conv1 module and conv2 module of the OSNet backbone network, and Process and obtain The common feature map of ;
[0093] The conv1 module uses the first 2D convolutional layer, the InstanceNorm2D normalization layer, and the first ReLU activation function to operate on Perform feature extraction and obtain The first stage common feature map ; In this embodiment, the number of input and output channels of the first 2D convolutional layer is (3, 64), the convolution kernel size is 7, and the step size is 2.
[0094] The conv2 module uses the MaxPool2D layer After global maximum pooling, we get The first stage of the common normalized feature map , and then use the first bottleneck module to Perform feature extraction and obtain The public feature map of the second stage .
[0095] Step 2.2, construct a global branch network, which is composed of the second and third bottleneck modules, the second 2D convolution layer, the first BatchNorm2d normalization layer, the second ReLU activation function, and the global feature module in sequence;
[0096] After being processed by the second and third bottleneck modules, we get The global branch common feature map , and then processed by the second 2D convolution layer, the first BatchNorm2D normalization layer and the second ReLU activation function, it is transformed into The global branch initial feature map ; In this embodiment, the number of input and output channels of the second 2D convolutional layer is (512, 512), the convolution kernel size is 1, and the step size is 1.
[0097] Global feature module pair Perform 2D average pooling to obtain The global branch feature map of , and remove After the first 1 / 3 of the most activated part, Perform point multiplication, and then perform 2D maximum pooling on the result to obtain The global branch robust feature map of .
[0098] Step 2.3, construct a local-based branch network, which is composed of the fourth and fifth bottleneck modules, the third 2D convolutional layer, the second BatchNorm2D normalization layer, the third ReLU activation function, and two branch local feature modules in sequence;
[0099] After being processed by the fourth and fifth bottleneck modules, we get The local branch common feature map , and then processed by the third 2D convolution layer, the second BatchNorm2D normalization layer and the third ReLU activation function, we get The local branch initial feature map ; In this embodiment, the number of input and output channels of the third 2D convolutional layer is (512, 512), the convolution kernel size is 1, and the step size is 1.
[0100] Will Input the first branch local feature module to perform 2D maximum pooling processing, and obtain The local branch global feature map ;
[0101] Will Input the second branch local feature module to perform adaptive 2D average pooling processing, and obtain The initial local feature map of the local branch In this embodiment, The shape of is 512×1×2; Divide it into two halves along the third dimension, and we get The first feature map of the local branch and The second feature map of the local branch .
[0102] Step 2.4, construct a channel-based branch network, which is composed of the sixth and seventh bottleneck modules, the fourth 2D convolutional layer, the third BatchNorm2D normalization layer, the fourth ReLU activation function and the channel feature module in sequence;
[0103] After being processed by the sixth and seventh bottleneck modules, we get Channel branch common feature map , Then, after the fourth 2D convolutional layer, the third BatchNorm2D normalization layer, and the fourth ReLU activation function, the result is subjected to 2D average pooling. Initial feature map of the channel branch ; In this embodiment, the number of input and output channels of the fourth 2D convolutional layer is (512, 512), the convolution kernel size is 1, and the step size is 1.
[0104] The channel feature module will The channel dimension is divided into two feature maps, and then the channel dimensions of the two feature maps are increased by 2D convolution with input and output channels of (256, 512) respectively, and the result is The first feature map of the channel branch and The second feature map of the channel branch .
[0105] In this embodiment, any bottleneck module in step 2.2, step 2.3, and step 2.4 is composed of a convolution block with a dimension of 1×1, an OS module, and a 2D average pooling layer in sequence; Figure 3 As shown;
[0106] Step 2.1.1: The 1×1 convolutional block in the first bottleneck module uses the fifth 2D convolutional layer, the fourth BatchNorm2D normalization layer, and the fifth ReLU activation function to operate on Perform nonlinear transformation to obtain The first stage nonlinear common feature map ;
[0107] Step 2.1.2: The OS module in the first bottleneck module uses the sixth 2D convolutional layer, the fifth BatchNorm2D normalization layer, and the sixth ReLU activation function to operate on Perform dimensionality reduction and obtain The first stage of dimensionality reduction common feature map ;
[0108] The OS module in the first bottleneck module uses the seventh 2D convolutional layer to Perform feature extraction and obtain The second stage first scale feature map .
[0109] The OS module in the first bottleneck module uses the eighth 2D convolutional layer, the ninth 2D convolutional layer in turn. Perform feature extraction and obtain The second stage second scale feature map ;
[0110] The OS module in the first bottleneck module uses the tenth 2D convolutional layer, the eleventh 2D convolutional layer, and the twelfth 2D convolutional layer in sequence. Perform feature extraction and obtain The second stage third scale feature map ;
[0111] The OS module in the first bottleneck module uses the thirteenth 2D convolution layer, the fourteenth 2D convolution layer, the fifteenth 2D convolution layer, and the sixteenth 2D convolution layer in sequence. Perform feature extraction and obtain The fourth scale feature map of the second stage ;
[0112] The 2D average pooling layer in the first bottleneck module is , , , The 2D average pooling and the seventh ReLU activation function are performed respectively, and the results are concatenated in the channel dimension to obtain The second stage multi-scale feature map .
[0113] In this embodiment, the number of input and output channels of the 1×1 convolution block in the first bottleneck module is (64, 64), the convolution kernel size is 1, and the stride is 1. The number of input and output channels of all other 2D convolution blocks is (64, 64), the convolution kernel size is 3, and the stride is 1.
[0114] The number of input and output channels of the 1×1 convolution blocks in the second, fourth, and sixth bottleneck modules are all (256, 96), the convolution kernel size is 1, and the stride is 1. The number of input and output channels of all other 2D convolution blocks are all (96, 96), the convolution kernel size is 3, and the stride is 1.
[0115] The number of input and output channels of the 1×1 convolution blocks in the third, fifth, and seventh bottleneck modules are all (384, 128), the convolution kernel size is 1, and the stride is 1. The number of input and output channels of all other 2D convolution blocks is (128, 128), the convolution kernel size is 3, and the stride is 1.
[0116] Step 2.5: Figure 4 As shown in Figure 2, a dimensionality reduction module is constructed, which consists of seven BNNeck modules and performs , , , , , , Processing to obtain multi-branch classification scores And multi-branch comprehensive feature map ;
[0117] The first BNNeck module pair Perform dimensionality reduction to obtain The global branch dimension reduction feature map , and then use the first fully connected layer to Process and obtain The global branch dimensionality reduction classification score ;
[0118] The second BNNeck module After dimensionality reduction, we get Global branch robust dimension reduction feature map , and then use the second fully connected layer to Process and obtain The global branch global dimensionality reduction classification score ;
[0119] The third BNNeck module is After dimensionality reduction, we get The local branch global dimension reduction feature map , and then use the third fully connected layer to Process and obtain The local branch robust dimensionality reduction classification score ;
[0120] The fourth BNNeck module pair After dimensionality reduction, we get The first dimension reduction feature map of the local branch , and then use the fourth fully connected layer to Process and obtain The global branch first dimension reduction classification score .
[0121] The fifth BNNeck module pair After dimensionality reduction, we get The second dimension reduction feature map of the local branch , and then use the fifth fully connected layer to Process and obtain The global branch second dimension reduction classification score ;
[0122] The sixth BNNeck module pair After dimensionality reduction, we get The first dimension reduction feature map of the channel branch , and then use the sixth fully connected layer to Process and obtain The first dimension reduction classification score of the channel branch ;
[0123] Seventh BNNeck module pair After dimensionality reduction, we get The second dimension reduction feature map of the channel branch , and then use the seventh fully connected layer to Process and obtain The second dimension reduction classification score of the channel branch .
[0124] Will , , , , , , Overlay Multi-branch classification score ;
[0125] Will , , Overlay Multi-branch comprehensive feature map .
[0126] Step 3, using formula (1) to construct the total loss function L of the open set electromagnetic radiation source individual feature extractor;
[0127] (1)
[0128] In formula (1), represents the cross entropy loss, represents the triple loss, represents the center loss, represents the circle loss, λ1, λ2, λ3, and λ4 are cross entropy losses respectively. , triplet loss , center loss The weight and circle loss weights; in this embodiment, λ1=0.5, λ2=1, λ3=0.0005, λ4=0.5.
[0129] Use formula (2) to construct the cross entropy loss :
[0130] (2)
[0131] In formula (2), x(k,i) is The symbolic function of If the true category prediction is correct, let x(k,i)=1, otherwise, let x(k,i)=0.
[0132] Use formula (3) to construct triple loss :
[0133] (3)
[0134] In formula (3), For Randomly selected from The same radio station Ultra-short wave frequency hopping IQ dual-channel signal Time-frequency diagram of The multi-branch comprehensive feature map, For Randomly selected from Not belonging to the same station Ultra-short wave frequency hopping IQ dual-channel signal Time-frequency diagram of Multi-branch comprehensive feature map; .
[0135] Use formula (4) to construct the center loss :
[0136] (4)
[0137] In formula (4), It is the mean center of the multi-branch comprehensive characteristic diagram of the time-frequency diagram of all ultra-short wave frequency-hopping IQ two-way signals belonging to the kth radio station.
[0138] Using formula (5) to construct the circle loss :
[0139] (5)
[0140] In formula (5), is the scale factor, is the inter-class weight factor of the circular loss of the kth station, is the intra-class weight factor of the circular loss of the kth station; in this embodiment, ; is the inter-class similarity of the circular loss of the kth radio station, is the intra-class similarity of the circle loss of the kth station, and:
[0141] (6)
[0142] (7)
[0143] In formula (6) and formula (7), for Middle The i-th ultra-short wave frequency-hopping IQ two-way signal of the radio station Time-frequency diagram of Multi-branch comprehensive feature map; for The i1th ultrashort wave frequency hopping IQ two-way signal of the kth station in Time-frequency diagram of Multi-branch comprehensive feature map; for The i2th ultrashort wave frequency hopping IQ two-way signal of the kth station Time-frequency diagram of Multi-branch comprehensive feature map.
[0144] Step 4: Use the ADAM optimizer to train the open set electromagnetic radiation source individual feature extractor until the total loss function L converges, thereby obtaining the optimal open set electromagnetic radiation source individual feature extractor; in this embodiment, during the training phase, the electromagnetic radiation source individual feature extractor is trained for 200 epochs, the batch size is set to 16, and the initial learning rate is set to ; After the 50th epoch, the learning rate decays to ; After the 100th epoch, the learning rate decays to ; After the 150th epoch, the learning rate decays to .
[0145] In this embodiment, 2400 time-frequency graphs obtained by wavelet transform of IQ dual-channel ultra-short wave frequency hopping signals of 12 radio stations in the same scene are used for training. During the training process, the size of the input time-frequency graph is randomly cropped to 384×384, and the total loss is calculated according to the deep features extracted by the electromagnetic radiation source individual feature extractor. The calculated loss is combined with the ADAM optimizer to guide the electromagnetic radiation source individual feature extractor for training, and convergence is obtained to obtain the optimal electromagnetic radiation source individual feature extractor;
[0146] Step 5: Use the open set electromagnetic radiation source individual feature extractor to retrieve the electromagnetic signal time-frequency diagram database dataset And the time-frequency diagram dataset of electromagnetic signals to be identified Processing is performed to obtain the corresponding retrieval library signal feature set And the signal feature set to be identified ;in, They are The global branch dimensionality reduction feature map, the global branch robust dimensionality reduction feature map, the local branch global dimensionality reduction feature map, the local branch first dimensionality reduction feature map, the local branch second dimensionality reduction feature map, the channel branch first dimensionality reduction feature map, the channel branch second dimensionality reduction feature map, They are The global branch dimensionality reduction feature map, the global branch robust dimensionality reduction feature map, the local branch global dimensionality reduction feature map, the local branch first dimensionality reduction feature map, the local branch second dimensionality reduction feature map, the channel branch first dimensionality reduction feature map, and the channel branch second dimensionality reduction feature map.
[0147] Will Overlay The encoding feature map of ;
[0148] Will Overlay The encoding feature map of ;
[0149] Use formula (8) to determine and Matching results :
[0150] (8)
[0151] In formula (8), d( ) is the cosine distance function, It is the station number obtained by extracting features and matching reasoning from the j2th IQ two-way signal to be identified of the k2th station.
[0152] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.
[0153] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.
[0154] Table 1 shows the ablation experiment comparison of the open set electromagnetic radiation source individual recognition method based on metric learning of the present invention with recognition accuracy as the evaluation index. For fair comparison, all open set electromagnetic radiation source individual recognition methods are retrained on the data set. During the test, all open set electromagnetic radiation source individual recognition methods use the wavelet transform RGB time-frequency diagram of the electromagnetic signal. The recognition accuracy is the proportion of the model's correct matching classification of samples of each category, and its value range is [0, 1]. The larger the value, the stronger the matching classification of the electromagnetic radiation source individual recognition method. According to the quantitative analysis of Table 1, it can be seen that the method of the present invention has achieved the best effect.
[0155] Table 1
[0156]
[0157] Among them, Ours represents the open set electromagnetic radiation source individual identification method based on metric learning of the present invention; RT represents the ResNet backbone network + Triplet Loss loss function method; RTC represents the ResNet backbone network + Triplet Loss loss function + Center Loss loss function method; RMTC represents the ResNet backbone network + multi-branch structure + Triplet Loss loss function + Center Loss loss function method; RMBTC represents the ResNet backbone network + multi-branch structure + BNNeck module + Triplet Loss loss function + Center Loss loss function method; RMBTCC represents the ResNet backbone network + multi-branch structure + BNNeck module + Triplet Loss loss function + Center Loss loss function + CircleLoss loss function method; OT represents the OSNet backbone network + Triplet Loss loss function method; OTC represents the OSNet backbone network + Triplet Loss loss function + Center Loss loss function method; OMTC represents the OSNet backbone network + multi-branch structure + Triplet Loss loss function + Center Loss loss function method; OMBTC stands for OSNet backbone network + multi-branch structure + BNNeck module + Triplet Loss loss function + Center Loss loss function method.
Claims
1. A method for identifying individual electromagnetic radiation sources based on open set metric learning, characterized in that: Follow these steps: Step 1: Obtain electromagnetic signal data sets of different radiation source individuals, preprocess them, and convert them into time-frequency diagrams; Step 1.1: Get the ultra-short wave frequency hopping signal set of K radio stations in the same outdoor scene ;in, represents the i-th ultra-short wave frequency-hopping IQ two-way signal of the k-th radio station in the same outdoor scene; represents the number of ultra-short wave frequency-hopping IQ two-way signals collected by the kth radio station in the same outdoor scene; At the same time, obtain The signal set of the ultra-short wave frequency hopping retrieval library of radio stations in the same outdoor scene and electromagnetic signal datasets to be identified ;in, It represents the j1th ultra-short wave frequency-hopping IQ two-way retrieval library signal of the k1th radio station in the same outdoor scene; represents the j2th IQ two-way signal to be identified of the k2th radio station in the same outdoor scene; It represents the number of ultra-short wave frequency-hopping IQ two-way retrieval library signals collected by the k1th radio station in the same outdoor scene; It indicates the number of ultra-short wave frequency-hopping IQ two-way signals to be identified collected by the k2th radio station in the same outdoor scene; ; Step 1.2: , and Perform wavelet transform and obtain the electromagnetic signal time-frequency diagram dataset accordingly , electromagnetic signal time-frequency diagram retrieval database data set And the time-frequency diagram dataset of electromagnetic signals to be identified ;in, express Time-frequency diagram obtained by wavelet transform; express Time-frequency diagram obtained by wavelet transform; express Time-frequency diagram obtained by wavelet transform; Step 2: Based on the multi-branch architecture, an open-set electromagnetic radiation source individual feature extractor is constructed, which includes: a common feature extraction module, a global-based branch network, a local-based branch network, a channel-based branch network, and a dimensionality reduction module; Step 2.1: Construct a common feature extraction module, which is composed of the conv1 module and conv2 module of the OSNet backbone network, and Process and obtain The common feature map of ; Step 2.2, construct a global branch network, which is composed of the second and third bottleneck modules, the second 2D convolution layer, the first BatchNorm2d normalization layer, the second ReLU activation function, and the global feature module in sequence; After being processed by the second and third bottleneck modules, we get The global branch common feature map , and then processed by the second 2D convolution layer, the first BatchNorm2D normalization layer and the second ReLU activation function, it is transformed into The global branch initial feature map ; Global feature module pair Perform 2D average pooling to obtain The global branch feature map of , and remove After the first 1 / 3 of the most activated part, Perform point multiplication, and then perform 2D maximum pooling on the result to obtain The global branch robust feature map of ; Step 2.3, construct a local-based branch network, which is composed of the fourth and fifth bottleneck modules, the third 2D convolutional layer, the second BatchNorm2D normalization layer, the third ReLU activation function, and two branch local feature modules in sequence; After being processed by the fourth and fifth bottleneck modules, we get The local branch common feature map , and then processed by the third 2D convolution layer, the second BatchNorm2D normalization layer and the third ReLU activation function, we get The local branch initial feature map ; Will Input the first branch local feature module to perform 2D maximum pooling processing, and obtain The local branch global feature map ; Will Input the second branch local feature module to perform adaptive 2D average pooling, and obtain Initial local feature map of the local branch ; then Divide it into two halves along the third dimension, and we get The first feature map of the local branch and The second feature map of the local branch ; Step 2.4, construct a channel-based branch network, which is composed of the sixth and seventh bottleneck modules, the fourth 2D convolutional layer, the third BatchNorm2D normalization layer, the fourth ReLU activation function and the channel feature module in sequence; After being processed by the sixth and seventh bottleneck modules, we get Channel branch common feature map , Then, after the fourth 2D convolution layer, the third BatchNorm2D normalization layer, and the fourth ReLU activation function, the result is subjected to 2D average pooling. Initial feature map of the channel branch ; The channel feature module will The channel dimension is divided into two feature maps, and then the channel dimensions of the two feature maps are increased by 2D convolution respectively. The first feature map of the channel branch and The second characteristic map of the channel branch ; Step 2.5: Construct a dimensionality reduction module, which consists of seven BNNeck modules and , , , , , , Processing to obtain multi-branch classification scores And multi-branch comprehensive feature map ; Step 3, using formula (1) to construct the total loss function L of the open set electromagnetic radiation source individual feature extractor; (1) In formula (1), represents the cross entropy loss, represents the triple loss, represents the center loss, represents the circle loss, λ1, λ2, λ3, and λ4 are cross entropy losses respectively. , triplet loss , center loss The weight and circle loss The weight of Step 4: using an ADAM optimizer to train the open set electromagnetic radiation source individual feature extractor until the total loss function L converges, thereby obtaining an optimal open set electromagnetic radiation source individual feature extractor; Step 5: Use the open set electromagnetic radiation source individual feature extractor to retrieve the electromagnetic signal time-frequency diagram database dataset And the time-frequency diagram dataset of electromagnetic signals to be identified Processing is performed to obtain the corresponding retrieval library signal feature set And the signal feature set to be identified ;in, They are The global branch dimensionality reduction feature map, the global branch robust dimensionality reduction feature map, the local branch global dimensionality reduction feature map, the local branch first dimensionality reduction feature map, the local branch second dimensionality reduction feature map, the channel branch first dimensionality reduction feature map, the channel branch second dimensionality reduction feature map, They are The global branch dimensionality reduction feature map, the global branch robust dimensionality reduction feature map, the local branch global dimensionality reduction feature map, the local branch first dimensionality reduction feature map, the local branch second dimensionality reduction feature map, the channel branch first dimensionality reduction feature map, and the channel branch second dimensionality reduction feature map; Will Overlay The encoding feature map of ; Will Overlay The encoding feature map of ; Use formula (8) to determine and Matching results : (8) In formula (8), d() is the cosine distance function, It is the station number obtained by extracting features and matching reasoning from the j2th IQ two-way signal to be identified of the k2th station.
2. The method for identifying individual open-set electromagnetic radiation sources based on metric learning according to claim 1, characterized in that: The conv1 module in step 2.1 sequentially uses the first 2D convolutional layer, the InstanceNorm2D normalization layer and the first ReLU activation function to operate on Perform feature extraction and obtain The first stage common feature map ; The conv2 module uses the MaxPool2D layer After global maximum pooling, we get The first stage common normalized feature map , and then use the first bottleneck module to Perform feature extraction and obtain The public feature map of the second stage .
3. The method for individual identification of open-set electromagnetic radiation sources based on metric learning according to claim 2, characterized in that: Any bottleneck module in step 2.2, step 2.3, and step 2.4 is composed of a convolution block with a dimension of 1×1, an OS module, and a 2D average pooling layer in sequence; Step 2.1.1: The 1×1 convolutional block in the first bottleneck module uses the fifth 2D convolutional layer, the fourth BatchNorm2D normalization layer, and the fifth ReLU activation function to operate on Perform nonlinear transformation to obtain The first stage nonlinear common feature map ; Step 2.1.2: The OS module in the first bottleneck module uses the sixth 2D convolutional layer, the fifth BatchNorm2D normalization layer, and the sixth ReLU activation function to operate on Perform dimensionality reduction and obtain The first stage of dimensionality reduction common feature map ; The OS module in the first bottleneck module uses the seventh 2D convolutional layer to Perform feature extraction and obtain The second stage first scale feature map ; The OS module in the first bottleneck module uses the eighth 2D convolutional layer, the ninth 2D convolutional layer in turn. Perform feature extraction and obtain The second stage second scale feature map ; The OS module in the first bottleneck module uses the tenth 2D convolutional layer, the eleventh 2D convolutional layer, and the twelfth 2D convolutional layer in sequence. Perform feature extraction and obtain The second stage third scale feature map ; The OS module in the first bottleneck module uses the thirteenth 2D convolution layer, the fourteenth 2D convolution layer, the fifteenth 2D convolution layer, and the sixteenth 2D convolution layer in sequence. Perform feature extraction and obtain The second stage fourth scale feature map ; The 2D average pooling layer in the first bottleneck module is , , , The 2D average pooling and the seventh ReLU activation function are performed respectively, and the results are concatenated in the channel dimension to obtain The second stage multi-scale feature map .
4. The method for identifying individual open-set electromagnetic radiation sources based on metric learning according to claim 3, characterized in that: The step 2.5 comprises: The first BNNeck module pair After dimensionality reduction, we get The global branch dimension reduction feature map , and then use the first fully connected layer to Process and obtain The global branch dimensionality reduction classification score ; The second BNNeck module After dimensionality reduction, we get Global branch robust dimension reduction feature map , and then use the second fully connected layer to Process and obtain The global branch global dimensionality reduction classification score ; The third BNNeck module is After dimensionality reduction, we get Local branch global dimension reduction feature map , and then use the third fully connected layer to Process and obtain The local branch robust dimensionality reduction classification score ; The fourth BNNeck module pair After dimensionality reduction, we get The first dimension reduction feature map of the local branch , and then use the fourth fully connected layer to Process and obtain The global branch first dimension reduction classification score ; The fifth BNNeck module pair After dimensionality reduction, we get The local branch second dimension reduction feature map , and then use the fifth fully connected layer to Process and obtain The global branch second dimension reduction classification score ; The sixth BNNeck module pair After dimensionality reduction, we get The first dimension reduction feature map of the channel branch , and then use the sixth fully connected layer to Process and obtain The first dimension reduction classification score of the channel branch ; Seventh BNNeck module pair After dimensionality reduction, we get The second dimension reduction feature map of the channel branch , and then use the seventh fully connected layer to Process and obtain The second dimension reduction classification score of the channel branch ; Will , , , , , , Overlay Multi-branch classification score ; Will , , Overlay Multi-branch comprehensive feature map .
5. The method for identifying individual open-set electromagnetic radiation sources based on metric learning according to claim 4, characterized in that: In step 3, the cross entropy loss is constructed using formula (2) : (2) In formula (2), x(k,i) is The symbolic function of If the true category prediction is correct, let x(k,i)=1, otherwise, let x(k,i)=0; Use formula (3) to construct the triple loss : (3) In formula (3), For Randomly selected from The same radio station Ultra-short wave frequency hopping IQ dual-channel signal Time-frequency diagram of The multi-branch comprehensive feature map, For Randomly selected from Not belonging to the same station Ultra-short wave frequency hopping IQ dual-channel signal Time-frequency diagram of Multi-branch comprehensive feature map; ; Use formula (4) to construct the center loss : (4) In formula (4), is the mean center of the multi-branch comprehensive characteristic diagram of the time-frequency diagram of all ultra-short wave frequency-hopping IQ two-way signals belonging to the k-th radio station; Using formula (5) to construct the circle loss : (5) In formula (5), is the scale factor, is the inter-class weight factor of the circular loss of the kth station, is the intra-class weight factor of the circular loss of the kth station; is the inter-class similarity of the circular loss of the kth radio station, is the intra-class similarity of the circle loss of the kth station, and: (6) (7) In formula (6) and formula (7), for Middle The i-th ultra-short wave frequency-hopping IQ two-way signal of the radio station Time-frequency diagram of Multi-branch comprehensive feature map; for The i1th ultrashort wave frequency hopping IQ two-way signal of the kth station in Time-frequency diagram of Multi-branch comprehensive feature map; for The i2th ultrashort wave frequency hopping IQ two-way signal of the kth station Time-frequency diagram of Multi-branch comprehensive feature map.
6. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the open-collection electromagnetic radiation source individual identification method described in any one of claims 1-5, and the processor is configured to execute the program stored in the memory.
7. 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 method for identifying an individual open-collection electromagnetic radiation source according to any one of claims 1 to 5 are executed.