Radiation source individual identification method based on inverse residual average pooling model

Abstract

The invention discloses a radiation source individual identification method based on an inverse residual average pooling model, which comprises the following steps of: firstly, constructing a pulse description word (PDW) parameter data set by extracting parameter characteristics of a radiation source, then constructing and initializing the inverse residual average pooling model, setting hyper-parameters and a training method of the model, and finally identifying the radiation source individual. Dividing the data set into a training set, a verification set and a test set, training the neural network model by using the training set, evaluating the training process by using the verification set, and after the training is completed, carrying out classification identification test on the data of the test set by using the trained model to realize individual identification and evaluate classification accuracy and model robustness. According to the method provided by the invention, the inverse residual convolution module and the global average pooling layer are combined, so that the feature information of the radiation source can be extracted more efficiently, the accuracy of individual identification of the radiation source is remarkably improved, the calculation complexity of network identification is effectively reduced, and the robustness of the network is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of radiation source individual identification, and in particular relates to a radiation source individual identification method based on an inverse residual average pooling model. Background Art

[0002] As the types and number of radiation sources continue to increase, the complexity of the space electromagnetic environment is also increasing, making it increasingly difficult to identify individual radiation sources. This has become a research hotspot in the field of signal processing.

[0003] Existing radiation source identification methods are based on manual identification, which achieves identification by extracting statistical features such as the mean, deviation and variance of the amplitude, frequency and phase parameters of the signal, or using the time domain, frequency domain and phase spectrum information of the original signal to obtain the individual radiation source's unique unintentional modulation information. However, with the rapid development of Internet of Things technology, the types and number of devices have increased dramatically, resulting in increased complexity, poor generalization and reduced recognition rate of existing methods, making it difficult to meet the requirements of large-scale data and high real-time performance. Therefore, scholars have proposed a radiation source identification algorithm based on neural networks to solve these problems.

[0004] In the research on individual identification of emitters based on neural networks, convolutional neural networks (CNN), recurrent neural networks (RNN, especially LSTM) and other network architectures or their combinations are mainly used. The document "Xu Xiong. Individual identification method of emitter targets using improved AlexNet. Telecommunication Engineering, 2018, 58(6)" directly uses the original time domain waveform image of the signal as input, combines multi-dimensional features, and uses a deep convolutional network similar to AlexNet for identification. The identification accuracy of 13 ADS-B emitters is 98.3%. The paper "PENG LN, ZHANG JQ, LIU M, et al. Deep Learning Based RF Fingerprint Identification Using Differential Constellation Trace Figure. IEEE Transactions on Vehicular Technology, 2020, 69(1): 1091-1095" extracts the RF fingerprint features of ZigBee devices through the differential constellation trace figure, and uses the mean clustering method for classification and identification. When the signal-to-noise ratio is 30dB and 15dB, the recognition accuracy of 54 ZigBee devices reaches 99.1% and 93.8% respectively.

[0005] However, the time-frequency domain image features or I / Q signal features used in the above studies often require complex data preprocessing when processing raw signals with high sampling rates. These preprocessing steps are not only time-consuming, but also generate a large amount of feature data, which greatly increases the computational complexity. In addition, the recognition ability of these network models is significantly reduced under low signal-to-noise ratio conditions, and their model robustness is poor. Summary of the invention

[0006] To solve the above technical problems, the present invention provides a method for individual identification of radiation sources based on an inverse residual average pooling model, which uses the pulse descriptor sequence of the radiation source signal as the individual identification feature of the radiation source, and uses a neural network model based on the inverse residual convolution architecture to perform individual identification of the radiation source, thereby improving its adaptive feature extraction capability and enhancing the robustness of the network model.

[0007] The technical solution adopted by the present invention is: a method for identifying individual radiation sources based on an inverse residual average pooling model, and the specific steps are as follows:

[0008] S1, constructing a pulse description word PDW parameter data set by extracting parameter features of the radiation source;

[0009] S2. Based on the inverse residual convolution module SandGlass Block and combined with the global average pooling layer, an initialized neural network model is constructed and parameters are initialized;

[0010] S3, setting the hyperparameters and training method of the neural network model in the training step S2;

[0011] S4, based on step S3, using the data set obtained in step S1 as the data source, and dividing it into a training set, a validation set and a test set according to a set ratio, using the training set data to train the neural network model initialized in step S2, and using the validation set to test the neural network during the training process;

[0012] The data set division ratio is set according to the actual situation. The training set, validation set and test set division ratio is set to a. 1 :a 2 :a 3 , then a 1 +a 2 +a 3 =1.

[0013] S5. Based on the neural network model obtained in step S4, classify and identify the unlabeled test set data described in step S4, and test the robustness of the model under the environment of leaky pulses and data measurement errors to obtain the individual identification result of the radiation source, complete the individual identification, and evaluate its classification accuracy and robustness.

[0014] Furthermore, the step S1 is specifically as follows:

[0015] Set the feature tensor size used for neural network model input to N s ×N p ×n, the input format of the PDW parameter dataset is a single-channel three-dimensional tensor.

[0016] Among them, the first dimension of the feature tensor is N s Represents the number of individual samples of the radiation source; the second dimension N of the feature tensor p Represents the number of types of pulse description word PDW parameters; the third dimension n of the feature tensor represents the number of time samplings of the reconnaissance frame, that is, each PDW parameter contains n measurement results.

[0017] The PDW parameters include: carrier frequency, arrival time, pulse amplitude, pulse width, pulse repetition frequency, arrival angle-pitch angle, arrival angle-azimuth angle, and intra-pulse modulation mode. The PDW parameter data set includes M types of different individual radiation sources, and the statistical distribution of each PDW parameter of each type of radiation source is as follows:

[0018] (1) Carrier frequency: M types of radiation sources are independent and identically distributed, and the carrier frequency in the radiation source sample has a RF ~b RF GHz is uniformly distributed, and the probability density function expression of the carrier frequency distribution is as follows:

[0019]

[0020] The carrier frequency has an amplitude of A RF The cosine unintentional modulation x[n] is expressed as follows:

[0021] x[n]=A RF ×rand×cos[2π(4i)x](i=1,2,3,...,M)

[0022] Among them, A RF represents the unintentional modulation amplitude of the carrier frequency, rand represents a random value in the range of 0 to 1, and i represents the individual number of the radiation source of different individuals in the data set, from 1 to M.

[0023] (2) Arrival time: M types of radiation sources are independent and identically distributed. The probability density function expression of the arrival time distribution in the radiation source sample is as follows:

[0024]

[0025] Among them, a TOA represents the minimum possible value of the uniform distribution of arrival time, b TOARepresents the maximum possible value of the uniform distribution for the arrival time.

[0026] (3) Pulse amplitude: M types of radiation sources are independent and identically distributed. The probability density function expression of the pulse amplitude distribution in the radiation source sample is as follows:

[0027]

[0028] where the pulse amplitude in the radiation source sample follows a Gaussian distribution, σ PA represents the standard deviation of the Gaussian distribution to which the pulse amplitude belongs, μ PA represents the mean of the Gaussian distribution to which the pulse amplitude belongs. The pulse amplitude has an amplitude of A PA The cosine unintentional modulation x[n] is expressed as follows:

[0029] x[n]=A PA ×rand×cos[2π(4i)x](i=1,2,3,...,M)

[0030] Among them, A PA Indicates that the pulse amplitude is not intended to modulate the amplitude.

[0031] (4) Pulse width: M types of radiation sources are independently distributed. The probability density function expression of the pulse width distribution in the radiation source sample is as follows:

[0032]

[0033] Among them, σ PW Indicates the standard deviation of the Gaussian distribution to which the pulse width belongs, μ PW represents the mean of the Gaussian distribution to which the pulse width belongs. The pulse width has an amplitude of A PW The cosine unintentional modulation x[n] is expressed as follows:

[0034] x[n]=A PW ×rand×cos[2π(4i)x](i=1,2,3,...,M)

[0035] Among them, A PW Indicates that the pulse width is not intended to modulate the amplitude.

[0036] (5) Pulse repetition frequency: The M types of radiation sources are independently distributed. The probability density function expression of the pulse repetition frequency distribution in the radiation source sample is as follows:

[0037]

[0038] where the pulse repetition frequency in the radiation source sample follows a Gaussian distribution, σ PRF represents the standard deviation of the Gaussian distribution to which the pulse repetition frequency belongs, μ PRFrepresents the mean of the Gaussian distribution to which the pulse repetition frequency belongs. The pulse repetition frequency has an amplitude of A PRF The cosine unintentional modulation x[n] is expressed as follows:

[0039] x[n]=A PRF ×rand×cos[2π(4i)x](i=1,2,3,...,M)

[0040] Among them, A PRF Indicates that the pulse repetition frequency does not unintentionally modulate the amplitude.

[0041] (6) Arrival angle-elevation angle: M types of radiation sources are independent and identically distributed. The probability density function of the arrival angle-elevation angle distribution in the radiation source sample is as follows:

[0042]

[0043] Among them, a DOA-El represents the minimum possible value of the uniform distribution of the arrival angle-pitch angle, b DOA-El Represents the maximum possible value of the uniform distribution of the angle of arrival-pitch angle.

[0044] (7) Arrival angle-azimuth: M types of radiation sources are independent and identically distributed. The probability density function of the arrival angle-azimuth distribution in the radiation source sample is as follows:

[0045]

[0046] Among them, a DOA-Az represents the minimum possible value of the uniform distribution of the angle-azimuth of arrival, b DOA-Az Represents the maximum possible value of the uniform distribution of the angle-azimuth of arrival.

[0047] (8) Intra-pulse modulation mode: M types of radiation sources are independent and identically distributed. The intra-pulse modulation mode in the radiation source sample is uniformly distributed in five ways. Let the random variable X represent the intra-pulse modulation mode of a radiation source. Its probability mass function PMF is expressed as follows:

[0048]

[0049] Among them, Sin means using high-frequency sine wave as the intra-pulse modulation method; LFM means using linear frequency modulation signal as the intra-pulse modulation method; Barker means using Barker phase coding as the intra-pulse coding method; 2PSK means using 2-phase coding as the intra-pulse coding method; 4PSK means using four-phase coding as the intra-pulse coding method.

[0050] Furthermore, the step S2 is specifically as follows:

[0051] The neural network model includes: 4 layers of SandGlass modules, 1 layer of global average pooling layer, 1 layer of linear fully connected layer, and 1 layer of data flattening layer.

[0052] Among them, each layer architecture of the neural network model is set in sequence, as follows:

[0053] 1) SandGlass module 1, including: point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, shortcut connection; the specific structure of this module is as follows:

[0054] a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows:

[0055] Convolutional layer, the number of input channels is 1, the number of output channels is 32, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0056] ReLU activation function layer;

[0057] Batch normalization layer with 32 input channels.

[0058] b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows:

[0059] Point-by-point convolution layer, the number of input channels is 1, the number of output channels is 4 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0060] ReLU activation layer;

[0061] Deep convolutional layer, the number of input channels is 4, the number of output channels is 4, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 4;

[0062] ReLU activation layer;

[0063] Point-wise convolution layer, the number of input channels is 4, the number of output channels is 32, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0064] ReLU activation layer;

[0065] Batch normalization layer with 32 input channels.

[0066] 2) SandGlass module 2, including: point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, and shortcut connection; the specific structure of this module is as follows:

[0067] a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows:

[0068] Convolutional layer: the number of input channels of the convolutional layer is 32, the number of output channels is 64, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0069] ReLU activation function layer;

[0070] Batch normalization layer with 64 input channels.

[0071] b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows:

[0072] Point-wise convolution layer, the number of input channels is 32, the number of output channels is 128 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0073] ReLU activation layer;

[0074] Deep convolutional layer, the number of input channels is 128, the number of output channels is 128, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 128;

[0075] ReLU activation layer;

[0076] Point-wise convolution layer, the number of input channels is 128, the number of output channels is 64, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0077] ReLU activation layer;

[0078] Batch normalization layer with 64 input channels.

[0079] 3) SandGlass module 3, including: point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, shortcut connection; the specific structure of this module is as follows:

[0080] a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows:

[0081] Convolutional layer: the number of input channels of the convolutional layer is 64, the number of output channels is 128, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0082] ReLU activation function layer;

[0083] Batch normalization layer with 128 input channels.

[0084] b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows:

[0085] Point-wise convolution layer, the number of input channels is 64, the number of output channels is 256 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0086] ReLU activation layer;

[0087] Deep convolutional layer, the number of input channels is 256, the number of output channels is 256, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 256;

[0088] ReLU activation layer;

[0089] Point-wise convolution layer, the number of input channels is 256, the number of output channels is 128, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0090] ReLU activation layer;

[0091] Batch normalization layer with 128 input channels.

[0092] 4) SandGlass module 4, including point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, and shortcut connection; the specific structure of this module is as follows:

[0093] a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows:

[0094] Convolutional layer: the number of input channels of the convolutional layer is 128, the number of output channels is 256, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0095] ReLU activation function layer;

[0096] Batch normalization layer with 256 input channels.

[0097] b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows:

[0098] Point-by-point convolution layer, the number of input channels is 128, the number of output channels is 512 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0099] ReLU activation layer;

[0100] Deep convolutional layer, the number of input channels is 512, the number of output channels is 512, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 512;

[0101] ReLU activation layer;

[0102] Point-wise convolution layer, the number of input channels is 512, the number of output channels is 256, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0103] ReLU activation layer;

[0104] Batch normalization layer with 256 input channels.

[0105] 5) Global average pooling layer with an output size of 1×1;

[0106] 6) Data flattening layer, nn.Flatten();

[0107] 7) Linear fully connected layer, the number of input neurons is 256 and the number of output neurons is 8.

[0108] Furthermore, the step S3 is specifically as follows:

[0109] Set the hyperparameters for training the neural network model, including: batch sample number N samples , network training learning rate lr, network training iteration number N epoch The value is determined as follows:

[0110] Assume that the loss function value of the neural network model after the kth iteration is L k . Set a threshold value ε th > 0 and a number of consecutive iterations N c If in N consecutive c During the iterations, the corresponding loss function reduction ratio is lower than the threshold value ε th , then the neural network model is considered to have reached convergence.

[0111] Define the ratio R of loss function reduction k The expression is as follows:

[0112]

[0113] The conditional expression for the network to converge is as follows:

[0114]

[0115] Where K = N epochRepresents the total number of iterations. The specific expression is as follows:

[0116]

[0117] If the above conditions are met, the neural network model is considered to have converged after the Kth iteration.

[0118] The training method of the neural network model is specifically as follows:

[0119] (1) Use cross entropy as the loss function between the network output vector and the standard label value;

[0120] For a sample, the cross entropy loss function is defined as follows:

[0121] Let y = (y 1 ,y 2 ,...,y C ) represents the true category distribution, Represents the predicted probability distribution of the neural network model.

[0122] Among them, C represents the number of individual categories of radiation sources, y i represents the true label of the i-th category, Represents the probability that the model predicts the i-th category.

[0123] The calculation expression of the cross entropy loss function MSELoss is as follows:

[0124]

[0125] When there are N samples in a batch, the average cross entropy loss function of the batch is The expression is defined as follows:

[0126]

[0127] Where N represents the number of samples in the batch, represents the true label of the αth sample, It represents the probability of the i-th class predicted by the neural network for the α-th sample.

[0128] (2) Using the Adam algorithm as an iterative optimizer to adjust the parameters of the neural network model;

[0129] 1) Initialization parameters;

[0130] The initialization parameters for the Adam iterative optimizer include:

[0131] a) learning rate lr;

[0132] b) Exponential decay rate β of the first-order moment estimate1 ;

[0133] c) Exponential decay rate β of the second-order moment estimate 2 ;

[0134] d) Minimum value ∈ that prevents division by zero;

[0135] e) time scale t=0 at the initialization moment, the initial value of the parameter θ;

[0136] f) The first-order moment estimation variable m and the second-order moment estimation variable v are initialized to 0, and the expressions are as follows:

[0137] m 0 =0,v 0 =0,t=0

[0138] 2) Calculate the gradient of the neural network model;

[0139] At each time step t, the gradient g of the above MSELoss loss function L(θ) with respect to the parameter θ is calculated t , the expression is as follows:

[0140]

[0141] Among them, g t represents the gradient at time step t, Represents the gradient of the loss function with respect to the parameters.

[0142] 3) Update the first-order moment estimate m t and the second-order moment estimate v t , the expression is as follows:

[0143] m t =β 1 m t-1 +(1-β 1 ) t

[0144] Among them, m t represents the first-order moment estimate at time step t, β 1 represents the exponential decay rate of the first-order moment estimate, g t represents the gradient at time step t.

[0145]

[0146] Among them, v t represents the second-order moment estimate at time step t, β 2 represents the exponential decay rate of the second-order moment estimate, Represents the square of the gradient.

[0147] 4) Deviation correction:

[0148] For m t and v t To correct the deviation, the expression is as follows:

[0149]

[0150] in, represents the bias-corrected first-order moment estimate, Represents β 1 to the power of t.

[0151]

[0152] in, represents the bias-corrected second-order moment estimate, Represents β 2 to the power of t.

[0153] 5) Update parameter θ:

[0154]

[0155] Among them, θ t+1 represents the updated parameters at time step t+1, represents the bias-corrected first-order moment estimate, represents the bias-corrected second-order moment estimate, and ∈ represents the minimum value to prevent division by zero.

[0156] Furthermore, in step S4, the neural network model initialized in step S2 is trained using the training set data, and the neural network is tested using the validation set during the training process, as follows:

[0157] A1. Load the divided training set, set the number of batch samples of the data, and generate a data iterator;

[0158] The number of batch samples of the training set data is the hyperparameter N in step S3. samples , and the data iterator shuffles the data in each iteration cycle.

[0159] A2. The training iteration starts, clearing the currently recorded neural network optimizer gradients for all parameters. θ L(θ t );

[0160] A3. Input the label data matrix and obtain the output data matrix;

[0161] For the input training set data matrix, first add a matrix dimension as the number of channels, and the initial number of channels is 1. After inputting the neural network, obtain the output result vector of the neural network, that is, the output data matrix.

[0162] A4. Calculate the loss function value;

[0163] The output vector of the neural network and the standard label vector are both 1×C vectors, where each value represents the possibility that the radiation source is the individual with that number, and the cross entropy function is used to calculate the model loss function.

[0164] A5. Calculate the recognition accuracy of the neural network on the validation set in the current iteration cycle through the validation set.

[0165] A6. Calculate the model gradient by performing neural network back propagation;

[0166] A7. Update the model parameters by calling the optimizer to complete parameter optimization.

[0167] Furthermore, the step S5 is specifically as follows:

[0168] S51, load the test set divided in step S4, set the number of batch samples of the data, and generate a data iterator for the test set;

[0169] The number of batch samples of the test set data is any value greater than 1 and is set according to the actual situation.

[0170] S52, testing is performed under the leakage pulse and data measurement error environment respectively, that is, the unlabeled test set feature vector is input into the neural network model trained in step S4 to obtain output, and the radiation source individual identification result is obtained;

[0171] S53, using accuracy to evaluate the correct prediction ratio between the neural network output result and the standard result, and calculating the average value of the classification and recognition accuracy of the neural network model for the entire test set samples as the test accuracy;

[0172] The trained model can predict the probability of each output category by inputting any sample features, and use the category with the highest predicted probability as the output category. If the prediction is consistent with the actual category, the prediction is correct. The performance of the model is evaluated using accuracy, which is equal to the ratio between the number of correct predictions and the total number of predictions.

[0173] Beneficial effects of the present invention: The method of the present invention first constructs a pulse description word parameter data set through the parameter characteristics of the radar radiation source, then constructs an initialized neural network model and performs parameter initialization, sets the hyperparameters and training methods of the neural network model, and then divides the data set into a training set, a validation set and a test set, uses the training set to train the neural network model, uses the validation set to test the neural network recognition accuracy during the training process, and finally uses the trained neural network model to perform classification and recognition tests on the test set data, completes radar individual recognition, and evaluates the classification accuracy of the neural network on the test set. The method of the present invention is based on the inverse residual convolution module SandGlass Block, combined with the global average pooling layer for classification and recognition, and can more effectively extract the characteristic information of the radar radiation source, thereby significantly improving the recognition accuracy and improving the adaptive feature extraction capability; the method of the present invention uses the pulse description word of the original signal of the radiation source as the input feature, which is smaller than the original signal data volume, and effectively reduces the computational complexity; the method of the present invention can effectively deal with the situation of missing data and enhance the robustness of the network by constructing a pulse description word data set and combining it with the neural network model training, even in the case of partial data missing, it can still maintain a high recognition accuracy, which is of great significance for the data incompleteness problem that may occur in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0174] Figure 1 The present invention is a flow chart of a method for identifying individual radiation sources based on an inverse residual average pooling model.

[0175] Figure 2 This is a graph showing the statistical distribution of the values ​​of each parameter in the data set constructed in the embodiment of the present invention.

[0176] Figure 3 This is a graph showing the change in the loss function during neural network training in an embodiment of the present invention.

[0177] Figure 4 This is a graph showing changes in individual recognition accuracy during neural network training in an embodiment of the present invention.

[0178] Figure 5 It is a schematic diagram of a curve showing a change in the leakage pulse rate and the classification accuracy of the test set data during the neural network test in an embodiment of the present invention.

[0179] Figure 6 This is a schematic diagram of a classification confusion matrix when the leakage pulse rate is 0.01 during the neural network test in an embodiment of the present invention.

[0180] Figure 7 This is a schematic diagram of a classification confusion matrix when the leakage pulse rate is 0.03 during the neural network test in an embodiment of the present invention.

[0181] Figure 8 Schematic diagram of the curve of PDW parameter error rate and classification accuracy during neural network testing in an embodiment of the present invention.

[0182] Fig. 9 This is a schematic diagram of a classification confusion matrix when the PDW parameter error rate is 0.01 during the neural network test in an embodiment of the present invention.

[0183] Fig.10 This is a schematic diagram of the classification confusion matrix when the PDW parameter error rate is 0.03 during the neural network test in an embodiment of the present invention. DETAILED DESCRIPTION

[0184] The method of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0185] like Figure 1 As shown, a flow chart of a radiation source individual identification method based on an inverse residual average pooling model of the present invention is shown, and the specific steps are as follows:

[0186] S1. Construct a pulse description word (PDW) parameter data set by extracting parameter features of the radiation source;

[0187] S2. Based on the inverse residual convolution module SandGlass Block and combined with the global average pooling layer, an initialized neural network model is constructed and parameters are initialized;

[0188] S3, setting the hyperparameters and training method of the neural network model in the training step S2;

[0189] S4, based on step S3, using the data set obtained in step S1 as the data source, and dividing it into a training set, a validation set and a test set according to a set ratio, using the training set data to train the neural network model initialized in step S2, and using the validation set to test the neural network during the training process;

[0190] The data set division ratio is set according to the actual situation. The training set, validation set and test set division ratio is set to a. 1 :a 2 :a 3 , then a 1 +a 2 +a 3 = 1. In this embodiment, the division ratio is 6:2:2.

[0191] S5. Based on the neural network model obtained in step S4, classify and identify the unlabeled test set data described in step S4, and test the robustness of the model under the environment of leaky pulses and data measurement errors to obtain the individual identification result of the radiation source, complete the individual identification, and evaluate its classification accuracy and robustness.

[0192] In this embodiment, the step S1 is specifically as follows:

[0193] Set the feature tensor size used for neural network model input to N s ×N p ×n, the input format of the PDW parameter dataset is a single-channel three-dimensional tensor.

[0194] Among them, the first dimension of the feature tensor is N s Represents the number of individual samples of the radiation source; the second dimension N of the feature tensor p represents the number of types of pulse description word PDW parameters; the third dimension n of the feature tensor represents the number of time sampling of the reconnaissance frame, that is, each PDW parameter contains n measurement results. In this embodiment, N s =50000, N p =8, n=400.

[0195] The PDW parameters include: carrier frequency, arrival time, pulse amplitude, pulse width, pulse repetition frequency, arrival angle-pitch angle, arrival angle-azimuth angle, and intra-pulse modulation mode.

[0196] The PDW parameter data set described in this embodiment includes M=8 types of radiation sources of different individuals. The definitions and related parameter distributions of the 8 types are the same as those of the training set, including 6342 groups of Class A, 6227 groups of Class B, 6327 groups of Class C, 6126 groups of Class D, 6310 groups of Class E, 6204 groups of Class F, 6181 groups of Class G, and 6283 groups of Class H.

[0197] In this embodiment, the statistical distribution results of the values ​​of each PDW parameter in the constructed data set are as follows: Figure 2 The PDW parameter data set includes M types of radiation sources of different individuals, and the statistical distribution of each PDW parameter of each type of radiation source is as follows:

[0198] (1) Carrier frequency: M types of radiation sources are independent and identically distributed, and the carrier frequency in the radiation source sample has a RF ~b RF GHz is uniformly distributed, and the probability density function expression of the carrier frequency distribution is as follows:

[0199]

[0200] The carrier frequency has an amplitude of A RFThe cosine unintentional modulation x[n] is expressed as follows:

[0201] x[n]=A RF ×rand×cos[2π(4i)x](i=1,2,3,...,M)

[0202] Among them, A RF represents the unintentional modulation amplitude of the carrier frequency, rand represents a random value in the range of 0 to 1, and i represents the individual number of the radiation source of different individuals in the data set, from 1 to M.

[0203] In this embodiment, the uniform distribution lower limit a RF =3GHz, uniform distribution upper limit b RF =5GHz, unintentional modulation amplitude A RF =0.2, number of radiation source types M = 8.

[0204] (2) Arrival time: M types of radiation sources are independent and identically distributed. The probability density function expression of the arrival time distribution in the radiation source sample is as follows:

[0205]

[0206] Among them, a TOA represents the minimum possible value of the uniform distribution of arrival time, b TOA Represents the maximum possible value of the uniform distribution for the arrival time.

[0207] In this embodiment, the number of radiation source types M = 8, the lower limit of the uniform distribution of arrival time a TOA =2×10 -2 ms, the upper limit of the uniform distribution of arrival time b TOA =6×10 -2 ms.

[0208] (3) Pulse amplitude: M types of radiation sources are independent and identically distributed. The probability density function expression of the pulse amplitude distribution in the radiation source sample is as follows:

[0209]

[0210] where the pulse amplitude in the radiation source sample follows a Gaussian distribution, σ PA represents the standard deviation of the Gaussian distribution to which the pulse amplitude belongs, μ PA represents the mean of the Gaussian distribution to which the pulse amplitude belongs. The pulse amplitude has an amplitude of A PA The cosine unintentional modulation x[n] is expressed as follows:

[0211] x[n]=A PA ×rand×cos[2π(4i)x](i=1,2,3,...,M)

[0212] Among them, A PA Indicates that the pulse amplitude is not intended to modulate the amplitude.

[0213] In this embodiment, the number of radiation source types M=8, the Gaussian distribution standard deviation σ PA =0.0625, Gaussian distribution mean μ PA =0.05×i+1.075(i=1,2,3,...,M), unintentional modulation amplitude A PA =0.05.

[0214] (4) Pulse width: M types of radiation sources are independently distributed. The probability density function expression of the pulse width distribution in the radiation source sample is as follows:

[0215]

[0216] Among them, σ PW Indicates the standard deviation of the Gaussian distribution to which the pulse width belongs, μ PW represents the mean of the Gaussian distribution to which the pulse width belongs. The pulse width has an amplitude of A PW The cosine unintentional modulation x[n] is expressed as follows:

[0217] x[n]=A PW ×rand×cos[2π(4i)x](i=1,2,3,...,M)

[0218] Among them, A PW Indicates that the pulse width is not intended to modulate the amplitude.

[0219] In this embodiment, the number of radiation source types M=8, the Gaussian distribution standard deviation σ PW =0.875, Gaussian distribution mean μ PW =0.35×i+5.35(i=1,2,3,...,M), unintentional modulation amplitude A PW =0.7.

[0220] (5) Pulse repetition frequency: The M types of radiation sources are independently distributed. The probability density function expression of the pulse repetition frequency distribution in the radiation source sample is as follows:

[0221]

[0222] where the pulse repetition frequency in the radiation source sample follows a Gaussian distribution, σ PRF represents the standard deviation of the Gaussian distribution to which the pulse repetition frequency belongs, μ PRF represents the mean of the Gaussian distribution to which the pulse repetition frequency belongs. The pulse repetition frequency has an amplitude of A PRF The cosine unintentional modulation x[n] is expressed as follows:

[0223] x[n]=A PRF ×rand×cos[2π(4i)x](i=1,2,3,...,M)

[0224] Among them, A PRF Indicates that the pulse repetition frequency does not unintentionally modulate the amplitude.

[0225] In this embodiment, the number of radiation source types M=8, the Gaussian distribution standard deviation σ PRF =(0.875×10 -2 ms) -1 , Gaussian distribution mean μ PRF =(0.35×10 -2 ×i+5.35×10 -2 ms) -1 (i=1,2,3,...,M), unintentional modulation amplitude A PRF =(0.7ms) -1 .

[0226] (6) Arrival angle-elevation angle: M types of radiation sources are independent and identically distributed. The probability density function of the arrival angle-elevation angle distribution in the radiation source sample is as follows:

[0227]

[0228] Among them, a DOA-El represents the minimum possible value of the uniform distribution of the arrival angle-pitch angle, b DOA-El Represents the maximum possible value of the uniform distribution of the angle of arrival-pitch angle.

[0229] In this embodiment, the lower limit a of the uniform distribution of the arrival angle and the pitch angle is DOA-El = -60°, arrival angle - upper limit of pitch angle uniform distribution b DOA-El =60°.

[0230] (7) Arrival angle-azimuth: M types of radiation sources are independent and identically distributed. The probability density function of the arrival angle-azimuth distribution in the radiation source sample is as follows:

[0231]

[0232] Among them, a DOA-Az represents the minimum possible value of the uniform distribution of the angle-azimuth of arrival, b DOA-Az Represents the maximum possible value of the uniform distribution of the angle-azimuth of arrival.

[0233] In this embodiment, the lower limit a of the uniform distribution of the angle of arrival and the azimuth angle is DOA-Az = 0°, the upper limit of the uniform distribution of the arrival angle-azimuth angle b DOA-Az =60°.

[0234] (8) Intra-pulse modulation mode: M types of radiation sources are independent and identically distributed. The intra-pulse modulation mode distribution in the radiation source sample is uniformly distributed in five ways. Let the random variable X represent the intra-pulse modulation mode of a radiation source, and its probability mass function (PMF) is:

[0235]

[0236] Among them, Sin means using high-frequency sine wave as the intra-pulse modulation method; LFM means using linear frequency modulation signal as the intra-pulse modulation method; Barker means using Barker phase coding as the intra-pulse coding method; 2PSK means using 2-phase coding as the intra-pulse coding method; 4PSK means using four-phase coding as the intra-pulse coding method.

[0237] Figure 2 The six parameters used by each radiation source are statistically analyzed in a histogram, where the horizontal axis of the histogram represents the specific value of the parameter, and the vertical axis of the histogram represents the number of entries of the parameter with the value in the total data generated by the radiation source.

[0238] In this embodiment, step S2 is specifically as follows:

[0239] The neural network model described in this embodiment includes: 4 layers of SandGlass modules, 1 layer of global average pooling layer, 1 layer of linear fully connected layer, and 1 layer of data flattening layer.

[0240] Among them, each layer architecture of the neural network model is set in sequence, as follows:

[0241] 1) SandGlass module 1, including: point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, shortcut connection; the specific structure of this module is as follows:

[0242] a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows:

[0243] Convolutional layer, the number of input channels is 1, the number of output channels is 32, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0244] ReLU activation function layer;

[0245] Batch normalization layer with 32 input channels.

[0246] b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows:

[0247] Point-by-point convolution layer, the number of input channels is 1, the number of output channels is 4 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0248] ReLU activation layer;

[0249] Deep convolutional layer, the number of input channels is 4, the number of output channels is 4, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 4;

[0250] ReLU activation layer;

[0251] Point-wise convolution layer, the number of input channels is 4, the number of output channels is 32, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0252] ReLU activation layer;

[0253] Batch normalization layer with 32 input channels.

[0254] 2) SandGlass module 2, including: point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, and shortcut connection; the specific structure of this module is as follows:

[0255] a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows:

[0256] Convolutional layer: the number of input channels of the convolutional layer is 32, the number of output channels is 64, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0257] ReLU activation function layer;

[0258] Batch normalization layer with 64 input channels.

[0259] b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows:

[0260] Point-wise convolution layer, the number of input channels is 32, the number of output channels is 128 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0261] ReLU activation layer;

[0262] Deep convolutional layer, the number of input channels is 128, the number of output channels is 128, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 128;

[0263] ReLU activation layer;

[0264] Point-wise convolution layer, the number of input channels is 128, the number of output channels is 64, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0265] ReLU activation layer;

[0266] Batch normalization layer with 64 input channels.

[0267] 3) SandGlass module 3, including: point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, shortcut connection; the specific structure of this module is as follows:

[0268] a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows:

[0269] Convolutional layer: the number of input channels of the convolutional layer is 64, the number of output channels is 128, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0270] ReLU activation function layer;

[0271] Batch normalization layer with 128 input channels.

[0272] b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows:

[0273] Point-wise convolution layer, the number of input channels is 64, the number of output channels is 256 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0274] ReLU activation layer;

[0275] Deep convolutional layer, the number of input channels is 256, the number of output channels is 256, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 256;

[0276] ReLU activation layer;

[0277] Point-wise convolution layer, the number of input channels is 256, the number of output channels is 128, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0278] ReLU activation layer;

[0279] Batch normalization layer with 128 input channels.

[0280] 4) SandGlass module 4, including point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, and shortcut connection; the specific structure of this module is as follows:

[0281] a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows:

[0282] Convolutional layer: the number of input channels of the convolutional layer is 128, the number of output channels is 256, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0283] ReLU activation function layer;

[0284] Batch normalization layer with 256 input channels.

[0285] b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows:

[0286] Point-by-point convolution layer, the number of input channels is 128, the number of output channels is 512 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0287] ReLU activation layer;

[0288] Deep convolutional layer, the number of input channels is 512, the number of output channels is 512, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 512;

[0289] ReLU activation layer;

[0290] Point-wise convolution layer, the number of input channels is 512, the number of output channels is 256, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0;

[0291] ReLU activation layer;

[0292] Batch normalization layer with 256 input channels.

[0293] 5) Global average pooling layer with an output size of 1×1;

[0294] 6) Data flattening layer, nn.Flatten();

[0295] 7) Linear fully connected layer, the number of input neurons is 256 and the number of output neurons is 8.

[0296] In this embodiment, step S3 is specifically as follows:

[0297] Set the hyperparameters for training the neural network model, including: batch sample number N samples =500, network training learning rate lr = 1×10 -5 , Network training iteration number N epoch The value is determined as follows:

[0298] Assume that the loss function value of the neural network model after the kth iteration is L k . Set a threshold value ε th > 0 and a number of consecutive iterations N c If in N consecutive c During the iterations, the corresponding loss function reduction ratio is lower than the threshold value ε th , then the neural network model is considered to have reached convergence.

[0299] Define the ratio R of loss function reduction k The expression is as follows:

[0300]

[0301] The conditional expression for the network to converge is as follows:

[0302]

[0303] Where K = N epoch Represents the total number of iterations. The specific expression is as follows:

[0304]

[0305] If the above conditions are met, the neural network model is considered to have converged after the Kth iteration.

[0306] The training method of the neural network model is specifically as follows:

[0307] (1) Use cross entropy as the loss function between the network output vector and the standard label value;

[0308] Cross-Entropy Loss is a commonly used loss function for classification problems in machine learning and deep learning. It measures the difference between two probability distributions and is usually used to evaluate the distance between the predicted distribution and the true distribution. For a sample, the cross-entropy loss function is defined as follows:

[0309] Let y = (y 1 ,y 2 ,...,y C ) represents the true category distribution, Represents the predicted probability distribution of the neural network model.

[0310] Among them, C represents the number of individual categories of radiation sources, y i represents the true label of the i-th category (one-hot encoding, only one component is 1 and the rest are 0), Represents the probability that the model predicts the i-th category.

[0311] The calculation expression of the cross entropy loss function MSELoss is as follows:

[0312]

[0313] When there are N samples in a batch, the average cross entropy loss function of the batch is The expression is defined as follows:

[0314]

[0315] Where N represents the number of samples in the batch, represents the true label of the αth sample, It represents the probability of the i-th class predicted by the neural network for the α-th sample.

[0316] (2) Using the Adam algorithm as an iterative optimizer to adjust the parameters of the neural network model;

[0317] 1) Initialization parameters;

[0318] The initialization parameters for the Adam iterative optimizer include:

[0319] a) Learning rate lr = 1 × 10 -5 ;

[0320] b) Exponential decay rate β of the first-order moment estimate 1 =0.9;

[0321] c) Exponential decay rate β of the second-order moment estimate 2 =0.999;

[0322] d) Minimum value to prevent division by zero ∈ = 1 × 10 -8 ;

[0323] e) time scale t=0 at the initialization moment, the initial value of the parameter θ;

[0324] f) The first-order moment estimation variable m and the second-order moment estimation variable v are initialized to 0, and the expressions are as follows:

[0325] m 0 =0,v 0 =0,t=0

[0326] 2) Calculate the gradient of the neural network model;

[0327] At each time step t, the gradient g of the above MSELoss loss function L(θ) with respect to the parameter θ is calculated t , the expression is as follows:

[0328]

[0329] Among them, g t represents the gradient at time step t, Represents the gradient of the loss function with respect to the parameters.

[0330] 3) Update the first-order moment estimate m t and the second-order moment estimate v t , the expression is as follows:

[0331] m t =β 1 m t-1 +(1-β 1 ) t

[0332] Among them, m t represents the first-order moment estimate at time step t, β 1 represents the exponential decay rate of the first-order moment estimate, g t represents the gradient at time step t.

[0333]

[0334] Among them, v t represents the second-order moment estimate at time step t, β 2 represents the exponential decay rate of the second-order moment estimate, Represents the square of the gradient.

[0335] 4) Deviation correction;

[0336] In order to reduce the deviation at the initial moment, m t and v t To correct the deviation, the expression is as follows:

[0337]

[0338] in, represents the bias-corrected first-order moment estimate, Represents β 1 to the power of t.

[0339]

[0340] in, represents the bias-corrected second-order moment estimate, Represents β 2 to the power of t.

[0341] 5) Update parameter θ;

[0342]

[0343] Among them, θ t+1 represents the updated parameters at time step t+1, represents the bias-corrected first-order moment estimate, represents the bias-corrected second-order moment estimate, and ∈ represents the minimum value to prevent division by zero.

[0344] In this embodiment, the data set used for the neural network uses 50,000 sets of randomly generated data, and is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The classification statistics of the data values ​​are as follows: Figure 2 shown.

[0345] Furthermore, in step S4, the neural network model initialized in step S2 is trained using the training set data, and the neural network is tested using the validation set during the training process, as follows:

[0346] A1. Load the divided training set, set the number of batch samples of the data, and generate a data iterator;

[0347] The number of batch samples of the training set data is the hyperparameter N in step S3. samples , and the data iterator shuffles the data in each iteration cycle.

[0348] A2. The training iteration starts, clearing the currently recorded neural network optimizer gradients for all parameters

[0349] A3. Input the label data matrix and obtain the output data matrix;

[0350] For the input training set data matrix, first add a matrix dimension as the number of channels, and the initial number of channels is 1. After inputting the neural network, obtain the output result vector of the neural network, that is, the output data matrix.

[0351] A4. Calculate the loss function value;

[0352] The output vector of the neural network and the standard label vector are both 1×C vectors, where each value represents the possibility that the radiation source is the individual with that number. The cross entropy function is used to calculate the model loss function, such as Figure 3 shown.

[0353] A5. Use the validation set to evaluate the recognition accuracy of the neural network on the validation set in the current iteration cycle, such as Figure 4 As shown;

[0354] A6. Calculate the model gradient by performing neural network back propagation;

[0355] A7. Update the model parameters by calling the optimizer to complete parameter optimization.

[0356] Furthermore, the step S5 is specifically as follows:

[0357] S51, load the test set divided in step S4, set the number of batch samples of the data, and generate a data iterator for the test set;

[0358] The number of batch samples of the test set data is any value greater than 1 and is set according to the actual situation.

[0359] S52, testing is performed under the leakage pulse and data measurement error environment respectively, that is, the unlabeled test set feature vector is input into the neural network model trained in step S4 to obtain output, and the radiation source individual identification result is obtained;

[0360] S53, using accuracy to evaluate the correct prediction ratio between the neural network output result and the standard result, and calculating the average value of the classification and recognition accuracy of the neural network model for the entire test set samples as the test accuracy;

[0361] The trained model can predict the probability of each output category by inputting any sample features, and use the category with the highest predicted probability as the output category. If the prediction is consistent with the actual category (label), the prediction is correct. Accuracy is used to evaluate the performance of the model, which is equal to the ratio between the number of correct predictions and the total number of predictions.

[0362] In this embodiment, for a test set with 10,000 groups of data, the feature array in the test set is assigned zeros, that is, the data is made to present a leaky pulse attribute according to a specified probability, and the impact of the leaky pulse on the recognition accuracy of various types of neural networks is evaluated. The leaky pulse rate and the classification accuracy curve of CNN, LSTM, and the method of the present invention ("MobileNetXt" in the figure) for the test set data are plotted as follows: Figure 5 As shown, the key nodes of the missed pulse rate and their corresponding recognition accuracy are listed in Table 1.

[0363] Table 1

[0364]

[0365] The comparison network CNN used in the method of the present invention is a deep convolutional network + Softmax classification neural network, and the LSTM is a long short-term memory network (LSTM, Long Short-Term Memory).

[0366] When the neural network architecture is CNN, LSTM, and the method of the present invention, the curves of the leakage pulse rate and the classification accuracy of the test set data are plotted as follows: Figure 5 At the same time, the classification confusion matrix when the leakage pulse rate is 0.01 and 0.03 is plotted as follows Figure 6 , Figure 7 shown.

[0367] In the test of PDW parameter error (i.e., "data noise"), noise is added to the feature array in the test set, that is, a section of Gaussian noise data is superimposed on the feature data according to the specified signal-to-noise ratio, and the impact of PDW parameter error on the recognition accuracy of various types of neural networks is evaluated. The data PDW parameter error rate (ratio of noise power to signal power) and the classification accuracy curve of CNN, LSTM, and the method of the present invention for the test set data are plotted as follows: Figure 8 As shown in Table 2, the influence of PDW parameter error rate on recognition accuracy is listed at intervals of 0.01.

[0368] Table 2

[0369]

[0370] When the neural network architecture is CNN, LSTM, and the method of the present invention, the PDW parameter error rate and the classification accuracy change curve of the test set data are plotted as follows: Figure 8 At the same time, the classification confusion matrix when the PDW parameter error rate is 0.01 and 0.03 is plotted as follows Fig. 9 , Fig.10 As shown. Figure 5 It can be seen that in the leakage pulse test, the classification accuracy of the three neural networks is from high to low: the method of the present invention>CNN>LSTM. Specifically, when the leakage pulse rate is 0.01, the classification accuracy of the present invention for the data is still 0.997, the classification accuracy of CNN is 0.989, and the classification accuracy of LSTM drops to 0.982.

[0371] On the other hand, from Figure 8 It can be seen that in the PDW parameter error test, the classification accuracy of the three neural networks is from high to low: the method of the present invention > CNN > LSTM. Specifically, when the PDW parameter error rate is 0.02, the classification accuracy of the present invention for the data is still 0.985, the classification accuracy of CNN is 0.951, and the classification accuracy of CLDNN drops to 0.886.

[0372] In summary, in this embodiment, the recognition accuracy of the method of the present invention in the task of identifying 8 types of radiation sources with a data scale of 10,000 is 99%. Under the conditions of a missing pulse rate of 0.03 and a parameter measurement error of 0.03, the recognition accuracy is 97%. It can be seen that the method of the present invention can not only improve the adaptive feature extraction capability, anti-missing and anti-noise capabilities of radiation source identification, effectively reduce the computational complexity of network identification, but also enhance the robustness of its network.

[0373] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. For those skilled in the art, the present invention may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A radiation source individual identification method based on the inverse residual average pooling model, the specific steps are as follows: S1, constructing a pulse description word PDW parameter data set by extracting parameter features of the radiation source; S2. Based on the inverse residual convolution module SandGlass Block and combined with the global average pooling layer, an initialized neural network model is constructed and parameters are initialized; S3, setting the hyperparameters and training method of the neural network model in the training step S2; S4, based on step S3, using the data set obtained in step S1 as the data source, and dividing it into a training set, a validation set and a test set according to a set ratio, using the training set data to train the neural network model initialized in step S2, and using the validation set to test the neural network during the training process; in, The data set division ratio is set according to the actual situation. If the training set, validation set and test set division ratio is a1:a2:a3, then a1+a2+a3=1; S5. Based on the neural network model obtained in step S4, classify and identify the unlabeled test set data described in step S4, and test the robustness of the model under the environment of leaky pulses and data measurement errors to obtain the individual identification result of the radiation source, complete the individual identification, and evaluate its classification accuracy and robustness.

2. According to claim 1, a radiation source individual identification method based on an inverse residual average pooling model is characterized in that: The step S1 is specifically as follows: Set the feature tensor size used for neural network model input to N s ×N p ×n, the input format of the PDW parameter data set is a single-channel three-dimensional tensor; Among them, the first dimension of the feature tensor is N s Represents the number of individual samples of the radiation source; the second dimension N of the feature tensor p Indicates the number of types of pulse description word PDW parameters; the third dimension n of the feature tensor represents the number of time samples of the reconnaissance frame, that is, each PDW parameter contains n measurement results; The PDW parameters include: carrier frequency, arrival time, pulse amplitude, pulse width, pulse repetition frequency, arrival angle-pitch angle, arrival angle-azimuth angle, and intra-pulse modulation mode; the PDW parameter data set includes M types of different individual radiation sources, and the statistical distribution of each PDW parameter of each type of radiation source is as follows: (1) Carrier frequency: M types of radiation sources are independent and identically distributed, and the carrier frequency in the radiation source sample has a RF ~b RF GHz is uniformly distributed, and the probability density function expression of the carrier frequency distribution is as follows: The carrier frequency has an amplitude of A RF The cosine unintentional modulation x[n] is expressed as follows: x[n]=A RF ×rand×cos[2π(4i)x] (i=1,2,3,...,M) Among them, A RF represents the unintentional modulation amplitude of the carrier frequency, rand represents a random value in the range of 0 to 1, and i represents the individual number of the radiation source of different individuals in the data set, from 1 to M; (2) Arrival time: M types of radiation sources are independent and identically distributed. The probability density function expression of the arrival time distribution in the radiation source sample is as follows: Among them, a TOA represents the minimum possible value of the uniform distribution of arrival time, b TOA represents the maximum possible value of the uniform distribution to which the arrival time belongs; (3) Pulse amplitude: M types of radiation sources are independent and identically distributed. The probability density function expression of the pulse amplitude distribution in the radiation source sample is as follows: where the pulse amplitude in the radiation source sample follows a Gaussian distribution, σ PA represents the standard deviation of the Gaussian distribution to which the pulse amplitude belongs, μ PA Represents the mean of the Gaussian distribution to which the pulse amplitude belongs; the pulse amplitude has an amplitude of A PA The cosine unintentional modulation x[n] is expressed as follows: x[n]=A PA ×rand×cos[2π(4i)x](i=1,2,3,...,M) Among them, A PA It indicates that the pulse amplitude has no intention to modulate the amplitude; (4) Pulse width: M types of radiation sources are independently distributed. The probability density function expression of the pulse width distribution in the radiation source sample is as follows: Among them, σ PW Indicates the standard deviation of the Gaussian distribution to which the pulse width belongs, μ PW Represents the mean of the Gaussian distribution to which the pulse width belongs; the pulse width has an amplitude of A PW The cosine unintentional modulation x[n] is expressed as follows: x[n]=A PW ×rand×cos[2π(4i)x](i=1,2,3,...,M) Among them, A PW This indicates that the pulse width does not unintentionally modulate the amplitude; (5) Pulse repetition frequency: The M types of radiation sources are independently distributed. The probability density function expression of the pulse repetition frequency distribution in the radiation source sample is as follows: where the pulse repetition frequency in the radiation source sample follows a Gaussian distribution, σ PRF represents the standard deviation of the Gaussian distribution to which the pulse repetition frequency belongs, μ PRF represents the mean of the Gaussian distribution to which the pulse repetition frequency belongs; the pulse repetition frequency has an amplitude of A PRF The cosine unintentional modulation x[n] is expressed as follows: x[n]=A PRF ×rand×cos[2π(4i)x](i=1,2,3,...,M) Among them, A PRF It indicates that the pulse repetition frequency unintentionally modulates the amplitude; (6) Arrival angle-elevation angle: M types of radiation sources are independent and identically distributed. The probability density function of the arrival angle-elevation angle distribution in the radiation source sample is as follows: Among them, a DOA-El represents the minimum possible value of the uniform distribution of the arrival angle-pitch angle, b DOA-El represents the maximum possible value of the uniform distribution of the arrival angle-pitch angle; (7) Arrival angle-azimuth: M types of radiation sources are independent and identically distributed. The probability density function of the arrival angle-azimuth distribution in the radiation source sample is as follows: Among them, a DOA-Az represents the minimum possible value of the uniform distribution of the angle-azimuth of arrival, b DOA-Az represents the maximum possible value of the uniform distribution of the angle-azimuth of arrival; (8) Intra-pulse modulation mode: M types of radiation sources are independent and identically distributed. The intra-pulse modulation modes in the radiation source samples are uniformly distributed in five ways. Let the random variable X represent the intra-pulse modulation mode of a radiation source. Its probability mass function PMF is expressed as follows: x i ∈{Sin,LFM,Barker,2PSK,4PSK} Among them, Sin means using high-frequency sine wave as the intra-pulse modulation method; LFM means using linear frequency modulation signal as the intra-pulse modulation method; Barker means using Barker phase coding as the intra-pulse coding method; 2PSK means using 2-phase coding as the intra-pulse coding method; 4PSK means using four-phase coding as the intra-pulse coding method.

3. According to claim 1, a radiation source individual identification method based on an inverse residual average pooling model is characterized in that: The step S2 is specifically as follows: The neural network model includes: 4 layers of SandGlass modules, 1 layer of global average pooling layer, 1 layer of linear fully connected layer, and 1 layer of data flattening layer; Among them, each layer architecture of the neural network model is set in sequence, as follows: 1) SandGlass module 1, including: point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, shortcut connection; the specific structure of this module is as follows: a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows: Convolutional layer, the number of input channels is 1, the number of output channels is 32, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation function layer; Batch normalization layer, the number of input channels is 32; b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows: Point-by-point convolution layer, the number of input channels is 1, the number of output channels is 4 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation layer; Deep convolutional layer, the number of input channels is 4, the number of output channels is 4, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 4; ReLU activation layer; Point-wise convolution layer, the number of input channels is 4, the number of output channels is 32, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation layer; Batch normalization layer, the number of input channels is 32; 2) SandGlass module 2, including: point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, and shortcut connection; the specific structure of this module is as follows: a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows: Convolutional layer: the number of input channels of the convolutional layer is 32, the number of output channels is 64, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation function layer; Batch normalization layer, the number of input channels is 64; b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows: Point-wise convolution layer, the number of input channels is 32, the number of output channels is 128 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation layer; Deep convolutional layer, the number of input channels is 128, the number of output channels is 128, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 128; ReLU activation layer; Point-wise convolution layer, the number of input channels is 128, the number of output channels is 64, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation layer; Batch normalization layer, the number of input channels is 64; 3) SandGlass module 3, including: point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, shortcut connection; the specific structure of this module is as follows: a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows: Convolutional layer: the number of input channels of the convolutional layer is 64, the number of output channels is 128, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation function layer; Batch normalization layer, the number of input channels is 128; b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows: Point-wise convolution layer, the number of input channels is 64, the number of output channels is 256 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation layer; Deep convolutional layer, the number of input channels is 256, the number of output channels is 256, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 256; ReLU activation layer; Point-wise convolution layer, the number of input channels is 256, the number of output channels is 128, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation layer; Batch normalization layer, the number of input channels is 128; 4) SandGlass module 4, including point-by-point convolution layer, batch normalization layer, ReLU activation layer, depth convolution layer, batch normalization layer, ReLU activation layer, point-by-point convolution layer, batch normalization layer, and shortcut connection; the specific structure of this module is as follows: a) If stride is 1 and in_channels is not equal to out_channels, the network hierarchy of the module is as follows: Convolutional layer: the number of input channels of the convolutional layer is 128, the number of output channels is 256, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation function layer; Batch normalization layer, the number of input channels is 256; b) If the input data does not meet the characteristics described in case a), the network hierarchy used is as follows: Point-by-point convolution layer, the number of input channels is 128, the number of output channels is 512 (in_channels*expansion), the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation layer; Deep convolutional layer, the number of input channels is 512, the number of output channels is 512, the convolution kernel size is 3×3, the stride is 2, the padding number is 1, and the number of groups is 512; ReLU activation layer; Point-wise convolution layer, the number of input channels is 512, the number of output channels is 256, the convolution kernel size is 1×1, the stride is 1, and the padding number is 0; ReLU activation layer; Batch normalization layer, the number of input channels is 256; 5) Global average pooling layer with an output size of 1×1; 6) Data flattening layer, nn.Flatten(); 7) Linear fully connected layer, the number of input neurons is 256 and the number of output neurons is 8.

4. According to claim 1, a radiation source individual identification method based on an inverse residual average pooling model is characterized in that: The step S3 is specifically as follows: Set the hyperparameters for training the neural network model, including: batch sample number N samples , network training learning rate lr, network training iteration number N epoch The value is determined as follows: Assume that the loss function value of the neural network model after the kth iteration is L k ; Set a threshold value ε th > 0 and a number of consecutive iterations N c ; If in N consecutive c During the iterations, the corresponding loss function reduction ratio is lower than the threshold value ε th , then the neural network model is considered to have reached convergence; Define the ratio R of loss function reduction k The expression is as follows: The conditional expression for the network to converge is as follows: R k <e th Where K = N epoch Represents the total number of iterations. The specific expression is as follows: If the above conditions are met, the neural network model is considered to have converged after the Kth iteration; The training method of the neural network model is specifically as follows: (1) Use cross entropy as the loss function between the network output vector and the standard label value; For a sample, the cross entropy loss function is defined as follows: Let y = (y1, y2, ..., y C ) represents the true category distribution, Represents the predicted probability distribution of the neural network model; Among them, C represents the number of individual categories of radiation sources, y i represents the true label of the i-th category, Represents the probability that the model predicts the i-th category; The calculation expression of the cross entropy loss function MSELoss is as follows: When there are N samples in a batch, the average cross entropy loss function of the batch is The expression is defined as follows: Where N represents the number of samples in the batch, represents the true label of the αth sample, It represents the probability of the i-th class predicted by the neural network for the α-th sample; (2) Using the Adam algorithm as an iterative optimizer to adjust the parameters of the neural network model; 1) Initialization parameters; The initialization parameters for the Adam iterative optimizer include: a) learning rate lr; b) exponential decay rate β1 of the first-order moment estimate; c) exponential decay rate β2 of the second-order moment estimate; d) Minimum value ∈ that prevents division by zero; e) time scale t=0 at the initialization moment, the initial value of the parameter θ; f) The first-order moment estimation variable m and the second-order moment estimation variable v are initialized to 0, and the expressions are as follows: m0=0,v0=0,t=0 2) Calculate the gradient of the neural network model; At each time step t, the gradient g of the above MSELoss loss function L(θ) with respect to the parameter θ is calculated t , the expression is as follows: g t =▽ θ L(θ t ) Among them, g t represents the gradient at time step t, ▽ θ L(θ t ) represents the gradient of the loss function with respect to the parameters; 3) Update the first-order moment estimate m t and the second-order moment estimate v t , the expression is as follows: m t =β1m t-1 +(1-β1)g t Among them, m t represents the first-order moment estimate at time step t, β1 represents the exponential decay rate of the first-order moment estimate, and g t represents the gradient at time step t; Among them, v t represents the second-order moment estimate at time step t, β2 represents the exponential decay rate of the second-order moment estimate, represents the square of the gradient; 4) Deviation correction: For m t and v t To correct the deviation, the expression is as follows: in, represents the bias-corrected first-order moment estimate, represents β1 to the power of t; in, represents the bias-corrected second-order moment estimate, represents β2 raised to the power of t; 5) Update parameter θ: Among them, θ t+1 represents the updated parameters at time step t+1, represents the bias-corrected first-order moment estimate, represents the bias-corrected second-order moment estimate, and ∈ represents the minimum value to prevent division by zero.

5. According to claim 1, a radiation source individual identification method based on an inverse residual average pooling model is characterized in that: In step S4, the neural network model initialized in step S2 is trained using the training set data, and the neural network is tested using the validation set during the training process, as follows: A1. Load the divided training set, set the number of batch samples of the data, and generate a data iterator; The number of batch samples of the training set data is the hyperparameter N in step S3. samples , and the data iterator shuffles the data in each iteration cycle; A2. The training iteration starts, clearing the currently recorded neural network optimizer gradients for all parameters. θ L(θ t ); A3. Input label data matrix and obtain output data matrix; For the input training set data matrix, first add a matrix dimension as the number of channels, and the initial number of channels is 1. After inputting the neural network, obtain the output result vector of the neural network, that is, the output data matrix; A4. Calculate the loss function value; The output vector of the neural network and the standard label vector are both 1×C vectors, where each value represents the possibility that the radiation source is the individual with that number, and the cross entropy function is used to calculate the model loss function; A5. Calculate the recognition accuracy of the neural network on the validation set in the current iteration cycle through the validation set. A6. Calculate the model gradient by performing neural network back propagation; A7. Update the model parameters by calling the optimizer to complete parameter optimization.

6. The method for identifying individual radiation sources based on the inverse residual average pooling model according to claim 1, characterized in that: The step S5 is specifically as follows: S51, load the test set divided in step S4, set the number of batch samples of the data, and generate a data iterator for the test set; The number of batch samples of the test set data is any value greater than 1, which is set according to the actual situation; S52, testing is performed under the leakage pulse and data measurement error environment respectively, that is, the unlabeled test set feature vector is input into the neural network model trained in step S4 to obtain output, and the radiation source individual identification result is obtained; S53, using accuracy to evaluate the correct prediction ratio between the neural network output result and the standard result, and calculating the average value of the classification and recognition accuracy of the neural network model for the entire test set samples as the test accuracy; The trained model can predict the probability of each output category by inputting any sample features, and use the category with the highest predicted probability as the output category; if the prediction is consistent with the actual category, the prediction is correct; the accuracy is used to evaluate the performance of the model, and the accuracy is equal to the ratio between the number of correct predictions and the total number of predictions.

Images