Radar radiation source individual identification method based on comparative learning and temperature parameter self-adaption

Abstract

The invention relates to a radar radiation source individual identification method and system based on comparative learning and temperature parameter self-adaption. The method comprises the following steps: constructing a to-be-trained comparative learning identification network, and training the comparative learning identification network based on a pre-training stage and a fine tuning stage to obtain a trained comparative learning identification network; training a contrast learning recognition network by using the sample after data enhancement and a cross entropy loss function with an adaptive temperature coefficient to obtain a pre-trained network, the pre-trained network comprising a coding layer, a projection head layer and an output layer; performing fine tuning on the pre-trained network by using a fine tuning model to obtain a final comparative learning recognition network, the final comparative learning recognition network comprising a coding layer and a linear classification layer; the data features of the radar radiation source data are extracted by using the coding layer, and the identified radar radiation source individuals are output in the linear classification layer, so that the problem that a large amount of label data is needed in the existing radar radiation source individual identification task can be solved.

Description

Technical Field

[0001] Embodiments of the present application relate to the field of radar emitter identification, and particularly to a method and system for radar emitter individual identification based on contrast learning and temperature parameter adaptation. Background Art

[0002] In recent years, with the continuous development and application of artificial intelligence technology, more and more technologies have been applied to the field of radar emitter individual identification. Radar emitter individual identification is to identify radar individuals by extracting individual features, which is a hot research direction in the field of electronic countermeasures. Therefore, radar emitter individual identification based on deep learning has become the focus. However, with the increasingly complex electromagnetic environment and the increasingly subtle individual differences, the recognition effect of traditional deep learning algorithms is not good. Currently, the commonly used methods for radar emitter individual identification are divided into two categories: one is supervised deep learning, and the other is unsupervised deep learning, while semi-supervised learning is between the two.

[0003] Supervised deep learning mainly relies on pseudo-labels and true labels to train the deep learning model. The model is trained through existing true labels to obtain the deep features of the data for data classification. However, the premise of this process is to obtain a large number of labeled data for model training, and it is even more difficult to obtain a large number of labeled samples in a complex electromagnetic environment, which undoubtedly greatly increases the time and labor costs.

[0004] Unsupervised deep learning algorithms do not require a large number of labeled data. They learn the deep features of individuals through deep learning networks, maximize the differences between individuals, and minimize the differences with themselves, so as to distinguish different samples to achieve the purpose of unsupervised individual identification. Traditional unsupervised radar emitter individual identification usually uses methods such as artificial feature extraction, such as time-frequency analysis, high-order spectrum analysis, etc., and then uses unsupervised algorithms such as clustering for individual identification. However, these artificial features have obvious limitations and cannot represent all the features of individuals, and the effect is often not good in practical applications. Summary of the Invention

[0005] In view of this, embodiments of the present application propose a method and system for radar emitter individual identification based on contrast learning and temperature parameter adaptation, aiming to overcome the defect that the traditional unsupervised algorithm model loses information and cannot extract effective individuals when processing radar emitter data for individual identification.

[0006] To achieve the above object, an embodiment of the present application provides a radar emitter individual recognition method based on contrast learning and temperature parameter adaptation, including: obtaining preprocessed radar emitter data; constructing a contrast learning recognition network to be trained, where the contrast learning recognition network is trained based on a pre-training stage and a fine-tuning stage; among them, in the pre-training stage, the contrast learning recognition network is trained using data-augmented samples and a cross-entropy loss function with an adaptive temperature coefficient to obtain a pre-trained network, and the pre-trained network includes an encoding layer, a projection head layer, and an output layer; in the fine-tuning stage, a fine-tuning model is constructed, and the pre-trained network is fine-tuned using the fine-tuning model to obtain a final contrast learning recognition network, and the final contrast learning recognition network includes an encoding layer and a linear classification layer; the data features of the radar emitter data are extracted using the encoding layer, and the recognized radar emitter individuals are output in the linear classification layer.

[0007] Optionally, the encoding layer includes: a convolutional layer, a batch normalization layer, a ReLU activation layer, a residual block layer, and an average pooling layer connected in sequence, with a residual connection between the batch normalization layer and the ReLU activation layer;

[0008] Among them, the expression of the convolutional layer is:

[0009] C1(x) = w * x + b

[0010] In the formula, w and b are the convolution kernel and the bias respectively;

[0011] The expression of the batch normalization layer is:

[0012]

[0013] In the formula, μ and σ 2 are the mean and variance respectively, and γ and β are the parameters to be learned;

[0014] The expression of the ReLU activation layer is:

[0015] R(x) = max(0, a)

[0016] In the formula, a is:

[0017] The expression of the residual connection is:

[0018] (x, h) = x + h

[0019] In the formula, h is the output after the convolutional layer and the batch normalization layer.

[0020] Optionally, the residual block layer adopts a ResNet1D structure, and the ResNet1D structure includes four residual block layers. Each of the residual block layers is composed of multiple BasicBlock1Ds, and the structural expression of each BasicBlock1D is:

[0021] out = R(B(x) * w1) * w2 + shortcut(x)

[0022] where w1 and w2 represent convolutional kernels, shortcut(x) represents a residual connection, and R(B(x) * w1) * w2 represents downsampling of x;

[0023] The expression of the average pooling layer is:

[0024]

[0025] where L is the sequence length.

[0026] Optionally, the projection head layer is used to map the output of the encoder to the target space. The projection head layer includes: a one-dimensional convolutional layer, a SiLU activation layer, a flattening layer, and a fully connected layer connected in sequence; where the expression of the SiLU activation layer is:

[0027]

[0028] In the formula, x represents the output after the batch normalization layer.

[0029] Optionally, the method further includes: constructing an NT-Xent loss function for the output layer, and scaling the dot product similarity of the NT-Xent loss function by using a temperature scaling parameter. The expression of the scaled NT-Xent loss function is:

[0030]

[0031] In the formula, s i,j represents the dot product similarity between two representations, τ represents the temperature scaling parameter, and s i,j / τ represents the dot product similarity s i ' j after temperature scaling; using the softmax function to process the dot product similarity s i ' j after temperature scaling to obtain a smoothed dot product similarity; constructing an optimized NT-Xent loss function according to the smoothed dot product similarity; where the temperature scaling parameter is obtained based on training.

[0032] Optionally, the process of determining the temperature scaling parameter includes: in the pre-training stage, calculate the training loss in the previous round of training of the contrastive learning recognition network, and determine the temperature scaling parameter in the current round of training based on the training loss in the previous round. Execute the above process until the training converges to obtain the final temperature scaling parameter, where the expression of the temperature scaling parameter is:

[0033] τ = 0.6 * (1 - e -kL ) + 0.2

[0034] In the formula, L represents the training loss in the previous round, and k represents a constant; perform a linear transformation on the temperature scaling parameter to obtain a temperature scaling parameter belonging to a preset interval, and determine based on the temperature scaling parameter belonging to the preset interval.

[0035] Optionally, the process of obtaining the data-augmented samples at least includes: obtaining a time series set, where not all elements of the time series set are 0; using the DropoutMask method to process the time series set for the training set to obtain a time series set with some elements set to 0; using the time jitter algorithm to process the time series set with some elements set to 0 to obtain a distorted time series; randomly selecting a translation amount, and using the randomly selected translation amount to superimpose the distorted time series to obtain a translated time series; sampling the normal distribution function to obtain a scaling factor, and performing amplitude scaling on the translated time series based on the scaling factor to obtain an amplitude-scaled time series; and for each element in the sequence, generating Gaussian noise and superimposing the Gaussian noise on the amplitude-scaled time series to obtain two types of data-augmented samples.

[0036] Optionally, the construction process of the fine-tuning model includes: configuring the projection head layer in the pre-trained network as a linear classifier to obtain the fine-tuning model.

[0037] Optionally, using the fine-tuning model to fine-tune the pre-trained network to obtain the final contrastive learning recognition network includes: feeding the data-augmented sample subset and the corresponding labels into the contrastive learning recognition network together for each round of training; when performing each round of training, freeze the parameters of the encoding layer and update the parameters of the linear classification layer.

[0038] Based on the above embodiments, the present application further proposes a radar emitter individual recognition system based on contrast learning and temperature parameter adaptation, including: an acquisition module, which acquires preprocessed radar emitter data; a network construction module, which constructs the contrast learning recognition network, and the contrast learning recognition network is trained based on a pre-training stage and a fine-tuning stage; wherein, in the pre-training stage, enhanced samples are input into the contrast learning recognition network to be trained, and the contrast learning recognition network to be trained includes an encoding layer, a projection head layer, and an output layer, and the pre-training network uses a loss function of temperature-scaled cross-entropy loss with an adaptive temperature coefficient; in the fine-tuning stage, a fine-tuning model is constructed, and the pre-training network is fine-tuned using the fine-tuning model to obtain the contrast learning recognition network; a processing module, which is used to extract the data features of the radar emitter data using the encoding layer and output the recognized radar emitter individual in the linear classification layer.

[0039] The radar emitter individual recognition method and system proposed in the embodiments of the present application include acquiring preprocessed radar emitter data; constructing a contrast learning recognition network to be trained, and the contrast learning recognition network is trained based on a pre-training stage and a fine-tuning stage; wherein, in the pre-training stage, the contrast learning recognition network is trained using enhanced samples and a cross-entropy loss function with an adaptive temperature coefficient to obtain a pre-training network, and the pre-training network includes an encoding layer, a projection head layer, and an output layer; in the fine-tuning stage, a fine-tuning model is constructed, and the pre-training network is fine-tuned using the fine-tuning model to obtain the final contrast learning recognition network, and the final contrast learning recognition network includes an encoding layer and a linear classification layer; the data features of the radar emitter data are extracted using the encoding layer and the recognized radar emitter individual is output in the linear classification layer, which can solve the problem of the need for a large amount of labeled data in the existing radar emitter individual recognition task. By introducing a contrast learning network and a temperature parameter adaptation mechanism, the generalization ability and classification performance of the model can be significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a flowchart of the radar emitter individual recognition method based on contrast learning and temperature parameter adaptation provided in an embodiment of the present application;

[0041] Figure 2 is a network structure diagram of the radar emitter individual recognition method based on contrast learning and temperature parameter adaptation provided in an embodiment of the present application;

[0042] Figure 3 is a structural block diagram of a radar emitter individual recognition device based on contrast learning and temperature parameter adaptation provided in an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the following will elaborate on each embodiment of this application in conjunction with the accompanying drawings. However, those of ordinary skill in the art can understand that in each embodiment of this application, many technical details are provided to help readers better understand this application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in this application can still be implemented. The division of the following embodiments is for convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined and cross-referenced with each other on the premise of not being contradictory.

[0044] In the existing Patent One, an unsupervised communication radiation source individual recognition method based on bispectrum feature contrast learning is proposed. Two parameter-sharing residual networks are used as the backbone network for feature contrast learning. The rectangular integral bispectrum feature is used to learn the feature representation, and then the extracted new feature representation is used to conduct contrast learning at the clustering cluster level to complete the classification and recognition task with a relatively high recognition accuracy.

[0045] The existing Patent Two proposes a radiation source recognition method, device, and storage medium based on adversarial training and deep clustering. First, the acquired radiation source signal is preprocessed to generate an image; an unsupervised CNN model is pre-trained using a network model optimizer and a training set to obtain a pre-trained unsupervised CNN model; the pre-trained unsupervised CNN model is re-trained using a clustering structure; a fine-tuned unsupervised CNN model is obtained; the target radiation source signal is input into the fine-tuned unsupervised CNN model to obtain the signal category of the radiation source signal. This method reduces manual participation, reduces errors and labor costs caused by human factors, and solves the problem of low recognition accuracy in traditional methods.

[0046] The existing Patent Three proposes a radio frequency fingerprint recognition method and system based on data augmentation and contrast learning. This method first collects radio frequency signals to obtain a training data set; according to the training data set, a specified number of samples are selected as a batch training data set; the outputs obtained by all samples through the convolutional neural network CNN are calculated; the average loss function value of each batch training process is calculated; the loss function gradient is calculated, and the network parameters are updated according to the gradient descent principle. By jointly optimizing the category sample consistency and label consistency in the training stage, the internal information of the samples and the correlation information between the original samples and the augmented samples can be effectively utilized. However, in the data preprocessing stage, the radiation source data is also converted into pictures for processing, and in addition, a simple CNN network also has poor performance when processing a large amount of unlabeled data.

[0047] A method for identifying individual radiation sources based on unsupervised momentum contrast learning is proposed in the existing Patent Four. This method uses a combined signal preprocessing method for data augmentation to enable the model to obtain more information. The current signal features are compared with the signal features in the dictionary, and a dynamic dictionary is used to enable it to learn the radiation source fingerprint features from the data again using prior knowledge. The InfoNCE contrast loss function is used. A fully connected layer is used to map the input of the model into a feature space of a fixed size to obtain a unified representation of the features. In the data preprocessing stage, classic image cropping, flipping, and noise addition operations are used, but converting the radiation source IQ signal into an image for data augmentation itself has an impact on the fingerprint features of the samples, especially for radar radiation source data.

[0048] Through the analysis of the existing technical solutions, it is found that most of the existing unsupervised learning complete networks adopt simple CNN network models and cannot effectively extract the deep individual features of radar radiation source data. In addition, when the existing contrast learning networks perform data preprocessing, they usually convert one-dimensional time series data into images for further processing. However, since the image processing method cannot fully capture the specific features of radar signals, such as phase information, polarization characteristics, etc., these features may be difficult to directly express in the image. Therefore, directly processing the one-dimensional time series radiation source data can better extract the individual features of the data to complete the classification task. In addition, for the loss function, the temperature parameter in the loss function based on temperature scaling needs to be determined through a large number of experimental comparisons in advance, which will greatly increase the time cost. The radar radiation source individual identification method based on contrast learning and temperature parameter adaptation can just solve these problems.

[0049] The contrastive learning network in this application is a self-supervised learning method that learns feature representations by comparing the similarities between data samples. Its core idea is to make the model pull closer similar samples (positive sample pairs) in the embedding space while pushing away dissimilar samples (negative sample pairs), so as to capture the internal structure of the data.

[0050] An embodiment of this application proposes an invention of a radar radiation source individual identification method based on contrast learning and temperature parameter adaptation. This method aims to solve the problem of the need for a large amount of labeled data in the existing radar radiation source individual identification task, and by introducing a contrast learning and temperature parameter adaptation mechanism, significantly improves the generalization ability and classification performance of the model.

[0051] The radar radiation source individual identification method based on contrast learning and temperature parameter adaptation may include:

[0052] S10. Obtain the preprocessed radar radiation source data;

[0053] S20. Construct a contrastive learning recognition network to be trained, where the contrastive learning recognition network is trained based on a pre-training stage and a fine-tuning stage. Specifically, in the pre-training stage, the contrastive learning recognition network is trained using data-augmented samples and a cross-entropy loss function with an adaptive temperature coefficient to obtain a pre-trained network, which includes an encoding layer, a projection head layer, and an output layer. In the fine-tuning stage, a fine-tuning model is constructed to fine-tune the pre-trained network to obtain the final contrastive learning recognition network, which includes an encoding layer and a linear classification layer.

[0054] S30. Use the encoding layer to extract the data features of the radar emitter data and output the recognized radar emitter individuals in the linear classification layer.

[0055] In the embodiment of the present application, the encoding layer includes:

[0056] A convolutional layer, a batch normalization layer, a ReLU activation layer, a residual block layer, and an average pooling layer connected in sequence, with a residual connection between the batch normalization layer and the ReLU activation layer.

[0057] Among them, the expression of the convolutional layer is:

[0058] C1(x) = w * x + b

[0059] In the formula, w and b are the convolution kernel and the bias respectively.

[0060] The expression of the batch normalization layer is:

[0061]

[0062] In the formula, μ and σ 2 are the mean and variance respectively, and γ and β are the parameters to be learned.

[0063] The expression of the ReLU activation layer is:

[0064] R(x) = max(0, a)

[0065] In the formula, a is:

[0066] The expression of the residual connection is:

[0067] (x, h) = x + h

[0068] In the formula, x is the output after passing through the convolutional layer and the batch normalization layer, and h is the value of the residual connection.

[0069] In the embodiment of the present application, the residual block layer adopts a ResNet1D structure, and the ResNet1D structure includes four residual block layers. Each residual block layer is composed of multiple BasicBlock1D, and the expression of each BasicBlock1D structure is:

[0070] out = R(B(x) * w1) * w2 + shortcut(x)

[0071] Among them, w1 and w2 represent convolutional kernels, shortcut(x) represents a residual connection, and R(B(x) * w1) * w2 represents downsampling of x;

[0072] The expression of the average pooling layer is:

[0073]

[0074] Among them, L is the sequence length.

[0075] In the embodiments of the present application, the projection head layer is used to map the output of the encoder to the target space. The projection head layer includes:

[0076] A one-dimensional convolutional layer, a SiLU activation layer, a flattening layer, and a fully connected layer connected in sequence;

[0077] Among them, the expression of the SiLU activation layer is:

[0078]

[0079] In the formula, x represents the output after the batch normalization layer.

[0080] In the embodiments of the present application, the method further includes:

[0081] Construct the NT-Xent loss function of the output layer, and scale the dot product similarity of the NT-Xent loss function using the temperature scaling parameter. The expression of the scaled NT-Xent loss function is:

[0082]

[0083] In the formula, s i,j represents the dot product similarity between two representations, τ represents the temperature scaling parameter, and s i,j / τ represents the dot product similarity s i ' j ;

[0084] Use the softmax function to process the dot product similarity s i ' j after temperature scaling to obtain the smoothed dot product similarity;

[0085] Construct the optimized NT-Xent loss function according to the smoothed dot product similarity;

[0086] Among them, the temperature scaling parameter is obtained based on training.

[0087] In an embodiment of the present application, the process of determining the temperature scaling parameter includes:

[0088] In the pre-training stage, calculate the training loss in the previous round of training of the contrastive learning recognition network, and determine the temperature scaling parameter in the current round of training based on the training loss in the previous round. Execute the above process until the training converges to obtain the final temperature scaling parameter. Among them, the expression of the temperature scaling parameter is:

[0089] τ = 0.6*(1 - e -kL ) + 0.2

[0090] In the formula, L represents the training loss in the previous round, and k represents a constant;

[0091] Perform a linear transformation on the temperature scaling parameter to obtain a temperature scaling parameter belonging to a preset interval, and determine based on the temperature scaling parameter belonging to the preset interval.

[0092] In an embodiment of the present application, the process of obtaining the data-augmented samples includes at least:

[0093] Obtain a time series set, where not all elements of the time series set are 0;

[0094] Use the DropoutMask method to process the time series set of the training set to obtain a time series set with some elements set to 0;

[0095] Use the time jitter algorithm to process the time series set with some elements set to 0 to obtain a distorted time series;

[0096] Randomly select a translation amount, and use the randomly selected translation amount to superimpose on the distorted time series to obtain a translated time series;

[0097] Sample the normal distribution function to obtain a scaling factor, and perform amplitude scaling on the translated time series based on the scaling factor to obtain an amplitude-scaled time series;

[0098] For each element S in the sequence i , generate Gaussian noise, and superimpose the Gaussian noise on the amplitude-scaled time series to obtain two types of data-augmented samples.

[0099] In an embodiment of the present application, constructing a fine-tuning model and using the fine-tuning model to fine-tune the pre-trained network to obtain the final contrastive learning recognition network includes:

[0100] Discard the projection head layer and the encoding layer of the pre-trained network, and add a classifier at the end of the encoding layer to obtain the fine-tuning model.

[0101] In an embodiment of the present application, the pre-trained network is fine-tuned using a fine-tuning model to obtain a contrastive learning recognition network, including:

[0102] The data-augmented sample subsets and corresponding labels are sent to the contrastive learning recognition network for each round of training. During each round of training, the encoding layer parameters are frozen and the linear classification layer parameters are updated. Each round of training includes a training mode and an evaluation mode. The training mode is used to fine-tune the model through data labels and network prediction labels. The evaluation mode is used to test the actual classification effect of the model, and return the loss value and recognition rate to determine the result of the fine-tuning.

[0103] On the basis of the above embodiments, the present application further provides a radar radiation source individual identification device based on contrast learning and temperature parameter adaptation. The radar radiation source individual identification device 100 includes:

[0104] The acquisition module 101 is used to acquire the pre-processed radar radiation source data;

[0105] The network construction module 102 is used to construct a contrastive learning recognition network, which is obtained through training in a pre-training stage and a fine-tuning stage; wherein, in the pre-training stage, the enhanced samples are input into the contrastive learning recognition network to be trained, wherein the contrastive learning recognition network to be trained includes a coding layer, a projection head layer, and an output layer, and the pre-trained network adopts a loss function of a temperature-scaled cross entropy loss with an adaptive temperature coefficient; in the fine-tuning stage, a fine-tuning model is constructed, and the pre-trained network is fine-tuned using the fine-tuning model to obtain the contrastive learning recognition network;

[0106] The processing module 103 is used to extract data features of radar emitter data using the coding layer, and output the identified radar emitter individuals in the linear classification layer.

[0107] Specifically, the above method embodiment may include the following execution process:

[0108] refer to Figure 2 , Figure 2 The network architecture of the radar emitter individual recognition method based on contrastive learning and temperature parameter adaptation is described in detail. The normalized raw data is input into the data enhancement module and the data enhancement operation is performed through the combined data enhancement method to obtain sample pairs, which are then input into the encoder network to obtain feature output. In the pre-training stage, the feature output is input into the projection head to obtain the projection vector. By maximizing the similarity, the distance between the same samples is shortened and the distance between different samples is increased to obtain the pre-trained network. The pre-trained network includes the encoding layer, the projection head layer and the output layer. In the model fine-tuning stage, the projection head is discarded and the downstream classifier is added at the end. A small amount of training set data is input into the network for fine-tuning, the encoder parameters are frozen, and only the downstream classifier is trained to obtain the classification network.

[0109] Step 1: Data preprocessing

[0110] Normalize all radar radiation source data. Subsequently, divide the processed data into a training set and a test set. The training set is used for model pre-training and fine-tuning, and the test set is used to evaluate the generalization performance of the model.

[0111] Step 2: Construct a pre-training network

[0112] The present invention designs a pre-training network based on contrastive learning, and its network structure is summarized as:

[0113] Input layer, encoding layer (ResNet1D), projection head layer Projector, and output layer (NT-Xent loss function)

[0114] Among them, the input layer receives one-dimensional time series data (x) with a dimension of ((N, C, L)), where: (N) is the batch size; (C) is the number of channels; (L) is the sequence length.

[0115] The following data augmentation methods are adopted to increase the diversity of samples:

[0116] Dropout method: For each element s in the time series S = {s1, s2,..., s n}, randomly decide whether to set it to zero with a certain probability and randomly discard some time series data points. i

[0117] Time jitter: For the time series S = {s1, s2,..., s n}, the original time steps T = {1, 2,..., n}, and the time warping steps are as follows:

[0118] For each time step t ∈ T, generate a random perturbation δ t ~N(0, σ 2 ). Calculate the warped time step t′ = t + δ t . Sort t ′ to obtain T′. Calculate the warped time series S′ by interpolation to randomly stretch or compress the time series.

[0119] Time translation: Perform a random sliding window operation on the time series. Specifically, for the time series S = {s1, s2,..., s n}, the time translation steps are as follows:

[0120] Randomly select a translation amount Δt from {-shift max , …, 0, …, shift max}. Calculate the translated time series S′ = S + S i+Δt mod n .

[0121] Amplitude scaling: Randomly scale the signal amplitude. Specifically, for the time series S = {s1, s2, …, s n}, the amplitude scaling steps are as follows:

[0122] Sample the scaling factor f from the normal distribution N(0, σ 2 ). Calculate the scaled time series S' i = S i × f.

[0123] Random deletion: Randomly delete some time series data. Exemplarily, for the time series S = {s1, s2, …, s n}, the steps for adding noise are as follows:

[0124] For each element Si in the sequence, generate Gaussian noise ò i ~N(μ, σ 2 ). Calculate the time series S' after adding noise i = S i + ò i

[0125] It should be noted that during the enhancement process, at least two methods need to be mixed to increase sample diversity.

[0126] Encoding layer (ResNet1D) In this application, a one-dimensional ResNet network is designed as the encoder to adapt to the radar emitter data of one-dimensional time series. The Resnet1D network consists of a convolutional layer, a batch normalization layer, a ReLU activation layer, a residual block layer, and an average pooling layer. There are a large number of individual features in the radar emitter data. However, the method of preprocessing the emitter data to generate pictures and then processing them will lose some individual features of the radar emitter. For example, the original emitter data will contain information such as time, frequency, and amplitude, but the time information will be lost after converting it into a picture. In addition, image processing methods cannot fully capture the specific features of radar signals, such as phase information and polarization characteristics, which may be difficult to directly express in images. Therefore, directly processing the one-dimensional time series of emitter data can better extract the individual features of the data to complete the classification task.

[0127] There is a large amount of individual characteristics in the radar radiation source data. However, the method of preprocessing the radar radiation source data to generate pictures and then processing them will lose some individual characteristics of the radar radiation source. For example, the original radar radiation source data will contain information such as time, frequency, and amplitude. However, after converting it into a picture, the time information will be lost. In addition, image processing methods cannot fully capture the specific characteristics of radar signals, such as phase information and polarization characteristics. These characteristics may be difficult to directly express in the image. Therefore, directly processing the one-dimensional time series of radar radiation source data can better extract the individual characteristics of the data and complete the classification task.

[0128] Specifically, its structure is as follows:

[0129] The convolutional layer is:

[0130] C1(x) = w * x + b

[0131] where w and b are the convolution kernel and bias respectively;

[0132] The batch normalization layer is:

[0133]

[0134] where μ and σ 2 are the mean and variance respectively, and γ and β are learnable parameters. The ReLU activation layer is:

[0135] R(x) = max(0, a)

[0136] The residual connection is:

[0137] (x, h) = x + h

[0138] where h is the output after convolution and batch normalization. ResNet1D contains four residual block layers (layer1 to layer4), and each layer is composed of multiple BasicBlock1D. The structure of each BasicBlock1D is as follows:

[0139] out = R(B(x) * w1) * w2 + shortcut(x)

[0140] where w1 and w2 are the convolution kernels, shortcut(x) is the residual connection, and R(B(x) * w1) * w2 is the result of downsampling x. The downsampling layer is used to adjust the dimension. The average pooling layer is:

[0141]

[0142] where L is the sequence length.

[0143] The ResNet1D network contains four residual block layers (layer1 to layer4), and each layer consists of multiple BasicBlock1D.

[0144] The projection head layer is used to map the output of the encoder to a low-dimensional space for calculating the contrastive loss. Its structure includes a one-dimensional convolutional layer, a SiLU activation layer, a flattening layer, and a fully connected layer.

[0145] Among them, the one-dimensional convolutional layer is used to extract features.

[0146] The expression of the SiLU activation layer is:

[0147]

[0148] The input data x passes through the one-dimensional convolutional layer to extract features. The output of the one-dimensional convolutional layer passes through the SiLU activation function to increase non-linearity.

[0149] The flattening layer is used to flatten the activated output, and the flattened vector passes through the fully connected layer to map to the target dimension and convert the features into a one-dimensional vector.

[0150] The fully connected layer is used to map to the target dimension. The output layer is the NT-Xent loss function.

[0151] The expression of the NT-Xent loss function is:

[0152]

[0153] Where s i,j is the dot product similarity between two representations. Temperature scaling is achieved by scaling the log function before the softmax function. The scaled similarity s i ' j The calculation formula is:

[0154]

[0155] Then, the softmax function is calculated as follows:

[0156]

[0157] When τ is large, the scaled similarity s i ' j is small, making the distribution of the softmax output smoother and the probability values closer to a uniform distribution. When τ is small, the scaled similarity s i ' j is large, making the distribution of the softmax output sharper and the probability values closer to a one-hot distribution.

[0158] For the loss function, the loss function based on temperature scaling largely depends on the selection of the temperature coefficient. Usually, a large number of experiments are required to determine its value in order to achieve better training results, which greatly increases the time cost. Therefore, the temperature coefficient is optimized and adaptively adjusted according to the training situation, especially the loss value.

[0159] Among them, the temperature coefficient τ is determined by the training loss of the previous round of the pre-trained network:

[0160] τ = 0.6 * (1 - e -kL ) + 0.2

[0161] Among them, k is a constant used to adjust the change rate, L is the training loss of the previous round of the pre-trained network. At the same time, to avoid the situation where the temperature coefficient approaches 1 or 0 when the training loss is too large or too small, during pre-training, the value of τ is stabilized between (0.2, 0.8) using a linear transformation.

[0162] In addition, when training the pre-trained network, the learning rate is calculated using cosine annealing:

[0163]

[0164] Among them, lr min and lr max represent the minimum and maximum learning rates, n represents the current training round, and N represents the total number of training rounds.

[0165] Step 3: Train the pre-trained network

[0166] When training the pre-trained network, the input parameters include the batch size, temperature coefficient, encoder function, and projection head function.

[0167] Secondly, for the sampled Perform the following operations:

[0168] 1: Extract two augmentation functions tT, t'T.

[0169] The first augmentation:

[0170]

[0171] z 2k-1 = g(h 2k-1 )

[0172] The second augmentation:

[0173]

[0174] z 2k = g(h 2k )

[0175] 2: Execute for all i ∈ {1, …, 2N} and j ∈ {1, …, 2N}:

[0176]

[0177] 3: Define the loss function l(i, j) as:

[0178]

[0179] 4: Calculate the overall loss L:

[0180]

[0181] 5: Update the networks f and g to minimize L.

[0182] Until the pre-training is completed to obtain the encoder network f(.) and the projection head g(.).

[0183] Step 4: Construct the fine-tuning model

[0184] Discard the projection head g(.) of the pre-trained model and only retain the encoder part for feature extraction. The specific structure is as follows:

[0185] The input data is subjected to feature extraction by the encoder (f), and the output feature vector (h) has a dimension of (N, D), where (N) is the batch size and (D) is the feature dimension. The feature vector is passed to the downstream classification task, and only the downstream classifier is trained.

[0186] Step 5: Model fine-tuning

[0187] Select a small amount of labeled training data and send its corresponding labels into the contrastive learning recognition network together. Freeze the encoder parameters and only update the newly added linear classification layer. Each round of training includes two modes, namely the training mode and the evaluation mode. The training mode includes fine-tuning the model through the data labels and the network prediction labels. The evaluation mode includes testing the true classification effect of the model and returning the loss value and the recognition rate.

[0188] Beneficial effects

[0189] The method proposed by the present invention has the following remarkable advantages:

[0190] No need for a large amount of labeled data: Through contrastive learning, only a small amount of labeled data is required to complete the model fine-tuning.

[0191] Adaptive temperature parameter: Dynamically adjust the temperature coefficient to effectively balance the similarity distribution of positive sample pairs.

[0192] High generalization ability: By combining pre-training and fine-tuning, the performance of the model in the individual recognition task of radar radiation sources is significantly improved.

[0193] The present invention provides an efficient and robust technical solution for the individual recognition of radar radiation sources.

[0194] The effects of the present invention will be further described below in conjunction with simulation experiments:

[0195] 1. Simulation conditions:

[0196] The hardware platform for the simulation experiment of the present invention is: the processor is Intel(R) Core i9-10980XE, the main frequency is 3.0GHZ, the memory is 64GB, and the graphics card is NVIDIA GeForce RTX 4090.

[0197] The software platform for the simulation experiment of the present invention is: WINDOWS10 operating system, MATLAB R2022a, Pytorch.

[0198] 2. Simulation content and result analysis:

[0199] The samples of radar radiation sources to be recognized used in the simulation experiment of the present invention are generated by MATLAB simulation. A total of 39,000 samples of 9 types of radar radiation source signals are generated. By changing the pulse width T, bandwidth B, carrier frequency fc of the radar radiation source, and the individual characteristic parameters of the radar radiation source, including carrier frequency offset, phase noise parameters, and the sampling frequency and cut-off frequency of the Butterworth filter, these 9 different individual signals of radar radiation sources are simulated. Among them, 30,000 samples form the training set, and 9,000 samples form the test set.

[0200] The individual recognition simulation experiment of the radar radiation source of the present invention uses the method of the present invention to perform individual recognition on 9 different individuals generated by simulation, and trains and tests the network model under different signal-to-noise ratios to verify the robustness of the model. The total number of correctly recognized samples in the test set data under each signal-to-noise ratio is counted, and then the total number of correctly recognized samples in the test set data under each signal-to-noise ratio is divided by the total number of samples in the test set data under each signal-to-noise ratio, which is 9,000, to obtain the correct recognition rate of radar individual recognition under each signal-to-noise ratio. All the calculation results are plotted in the following table.

[0201] The contrastive learning and temperature parameter adaptive network model is trained and tested under different signal-to-noise ratios to verify the robustness of the model. Since only the downstream classifier is trained in the model fine-tuning stage and the training speed is relatively fast, the number of stable iteration times of the correct recognition rate under different signal-to-noise ratios refers to the number of pre-training iterations. Four datasets with signal-to-noise ratios of 5dB, 10dB, 15dB, and 20dB are selected, and the experimental results are as follows:

[0202] Table 1 Recognition correct rate under different signal-to-noise ratios

[0203]

[0204] Table 2 Number of Stable Iterations of Correct Rate under Different Signal-to-Noise Ratios

[0205]

[0206] As can be seen from the above table, it can be observed that under the above signal-to-noise ratio conditions, as the number of pre-training times increases, the recognition accuracy can finally stabilize to near 100%, which shows the effectiveness of the one-dimensional Resnet plus projection head network; secondly, it can be found from Table 2 that when the signal-to-noise ratio is 5 dB, about 450 iterations are required for the recognition correct rate to stabilize, while when the signal-to-noise ratio is 10 dB, more than 375 iterations are required, and when the signal-to-noise ratio is increased to 20 dB, only about 200 iterations are required for the correct rate to stabilize. It can be seen that as the signal-to-noise ratio increases, the number of iterations required to reach the stable recognition rate also decreases accordingly. This indicates that as the signal conditions improve, the network model can fully learn the laws between the signal characteristics of different individuals within a shorter training time, and the weight coefficients of the one-dimensional Resnet plus projection head network model hardly need to be adjusted, and the recognition effect becomes more and more stable.

[0207] Table 3 Statistical Table of Recognition Correct Rates Obtained by Different Data Augmentation Methods

[0208]

[0209] As can be seen from the above table, the recognition correct rates obtained by all single data augmentation methods are generally low, while better results are achieved by using combined data augmentation methods. Therefore, it can be concluded that the composition of data augmentation operations is crucial for learning good feature representations. It can be speculated that when only the Dropout method is used, there may be a serious problem that the most important rising edge part of the data may be randomly discarded, and the neural network can use this problem to solve the prediction task, resulting in a very low final correct recognition rate. At the same time, more reasonable selection of the combination of data augmentation methods should also be noted.

[0210] Those of ordinary skill in the art can understand that the above embodiments are specific embodiments for implementing the present application, and in practical applications, various changes can be made in form and details without departing from the spirit and scope of the present application.

Claims

1. A radar radiation source individual identification method based on contrast learning and temperature parameter adaptation, characterized in that: include: Obtain preprocessed radar radiation source data; Constructing a contrastive learning recognition network to be trained, and training the contrastive learning recognition network based on the pre-training stage and the fine-tuning stage to obtain a trained contrastive learning recognition network; In the pre-training stage, the contrastive learning recognition network is trained using the data-enhanced samples and the cross entropy loss function with an adaptive temperature coefficient to obtain a pre-trained network, which includes a coding layer, a projection head layer and an output layer; In the fine-tuning stage, a fine-tuning model is constructed, and the pre-trained network is fine-tuned using the fine-tuning model to obtain a final contrastive learning recognition network, wherein the final contrastive learning recognition network includes an encoding layer and a linear classification layer; The data features of the radar emitter data are extracted using the encoding layer, and the identified radar emitter individuals are output in the linear classification layer.

2. The radar radiation source individual identification method based on contrast learning and temperature parameter adaptation according to claim 1 is characterized in that: The coding layer comprises: A convolutional layer, a batch normalization layer, a ReLU activation layer, a residual block layer and an average pooling layer connected in sequence, wherein the batch normalization layer and the ReLU activation layer are connected with a residual connection; Among them, the expression of the convolutional layer is: C1(x)=w*x+b Where w and b represent the convolution kernel and bias respectively; The expression of the batch normalization layer is: In the formula, μ and σ 2 denote the mean and variance respectively, γ, ò and β denote the parameters to be learned; The expression of the ReLU activation layer is: R(x)=max(0,a) In the formula, a represents: The expression of the residual connection is: (x,h)=x+h Wherein, h represents the output after the convolution layer and the batch normalization layer.

3. The radar radiation source individual identification method based on contrast learning and temperature parameter adaptation according to claim 2 is characterized in that: The residual block layer adopts the ResNet1D structure, and the ResNet1D structure includes four residual block layers, each of which is composed of multiple BasicBlock1Ds, and the structural expression of each BasicBlock1D is: out=R(B(x)*w1)*w2+shortcut(x) Among them, w1 and w2 represent convolution kernels, shortcut(x) represents residual connection, and R(B(x)*w1)*w2 represents downsampling of x; The expression of the average pooling layer is: Where L represents the sequence length.

4. The radar radiation source individual identification method based on contrast learning and temperature parameter adaptation according to claim 2 is characterized in that: The projection head layer is used to map the output of the encoder to the target space, and the projection head layer includes: The one-dimensional convolutional layer, SiLU activation layer, flattening layer and fully connected layer structure are connected in sequence; Wherein, the expression of the SiLU activation layer is: Wherein, x represents the output after the batch normalization layer.

5. The radar radiation source individual identification method based on contrast learning and temperature parameter adaptation according to claim 1 is characterized in that: The method further comprises: The NT-Xent loss function of the output layer is constructed, and the dot product similarity of the NT-Xent loss function is scaled using the temperature scaling parameter. The expression of the scaled NT-Xent loss function is: In the formula, s i,j represents the dot product similarity between two representations, τ represents the temperature scaling parameter, and s i,j / τ represents the dot product similarity s after temperature scaling i ' j ; Use the softmax function to scale the temperature dot product similarity s i ' j Processing is performed to obtain the smoothed dot product similarity; Constructing the optimized NT-Xent loss function according to the smoothed dot product similarity; Wherein, the temperature scaling parameter is obtained based on training.

6. The radar radiation source individual identification method based on contrast learning and temperature parameter adaptation according to claim 5 is characterized in that: The process of determining the temperature scaling parameter includes: In the pre-training stage, the training loss in the previous round of training on the contrastive learning recognition network is calculated, and the temperature scaling parameter in the current round of training is determined based on the training loss in the previous round. The above process is performed until the training converges to obtain the final temperature scaling parameter, wherein the expression of the temperature scaling parameter is: τ=0.6*(1-e -kL )+0.2 In the formula, L represents the training loss of the previous round, and k represents a constant; The final temperature scaling parameter is linearly transformed to obtain a temperature scaling parameter belonging to a preset interval.

7. The radar radiation source individual identification method based on contrast learning and temperature parameter adaptation according to claim 1 is characterized in that: The process of obtaining data-enhanced samples includes at least: Obtain a time series set, wherein the elements of the time series set are not all 0; The DropoutMask method is used to process the time series set of the training set to obtain a time series set with some elements set to 0; Processing the time series set where some elements are set to 0 using a time jitter algorithm to obtain a distorted time series; Randomly selecting a translation amount, and using the randomly selected translation amount to superimpose the distorted time series to obtain a translated time series; Sampling a normal distribution function to obtain a scaling factor, and performing amplitude scaling on the translated time series based on the scaling factor to obtain an amplitude-scaled time series; And, for each element in the sequence, generate Gaussian noise, and superimpose the Gaussian noise on the amplitude-scaled time series to obtain two types of data-enhanced samples.

8. The radar radiation source individual identification method based on contrast learning and temperature parameter adaptation according to claim 1 is characterized in that: The process of building a fine-tuning model includes; The projection head layer and the output layer in the pre-trained network are configured as linear classifiers to obtain a fine-tuning model.

9. The radar radiation source individual identification method based on contrast learning and temperature parameter adaptation according to claim 8, characterized in that: The method of fine-tuning the pre-trained network using the fine-tuning model to obtain a final contrastive learning recognition network includes: The data-enhanced sample subset and the corresponding labels are sent to the contrastive learning recognition network for each round of training; During each round of training, the encoding layer parameters are frozen and the linear classification layer parameters are updated.

10. A radar radiation source individual identification system based on contrast learning and temperature parameter adaptation, characterized in that: include: An acquisition module is used to acquire the pre-processed radar radiation source data; A network construction module is used to construct a contrastive learning recognition network to be trained, and to train the contrastive learning recognition network based on a pre-training stage and a fine-tuning stage to obtain a trained contrastive learning recognition network; In the pre-training stage, the contrastive learning recognition network is trained using the data-enhanced samples and the cross entropy loss function with an adaptive temperature coefficient to obtain a pre-trained network, which includes a coding layer, a projection head layer and an output layer; In the fine-tuning stage, a fine-tuning model is constructed, and the pre-trained network is fine-tuned using the fine-tuning model to obtain a final contrastive learning recognition network, wherein the final contrastive learning recognition network includes an encoding layer and a linear classification layer; The processing module is used to extract the data features of the radar emitter data by using the encoding layer, and output the identified radar emitter individuals in the linear classification layer.

Images