Method for identifying individual radiation sources based on class activation map and SincNet network

Abstract

The invention relates to a radiation source individual identification method based on a class activation map and a SincNet network. The purpose of the invention is to solve the problem that the existing pre-distortion technology weakens the nonlinear characteristics of the power amplifier of the radiation source, thereby weakening the radiation source individual identification performance. It is judged whether the radiation source signal to be measured is a high signal-to-noise ratio radiation source signal or a low signal-to-noise ratio radiation source signal. If the radiation source signal to be measured is a high signal-to-noise ratio radiation source signal, A to D are executed; if the radiation source signal to be measured is a low signal-to-noise ratio radiation source signal, one to three are executed; A, the power amplifier outputs a labeled signal; B: extracts a labeled CWD time-frequency distribution feature map; C: obtains a trained ResNet50 network; D: completes the radiation source individual signal after the pre-distortion to be measured; one: obtains a trained digital pre-distortion trainer; two: obtains a trained SincNet network; three: completes the identification of the radiation source individual signal after the pre-distortion to be measured. The present invention is used in the field of radiation source individual identification.

Description

Technical Field

[0001] The present invention relates to a method for identifying individual radiation sources. Background Art

[0002] Specific Emitter Identification (SEI) technology mainly realizes the identification and differentiation of radiation sources based on the unintentional modulation characteristics of signals. Power amplifier nonlinear distortion is one of the unintentional modulation characteristics and also one of the main bases for identifying individual radiation sources. However, in communication and radar systems, power amplifier nonlinear distortion leads to a decrease in the linear range and efficiency. With the expansion of the requirements for broadband communication and radar systems, power amplifier linearization technologies such as predistortion have emerged. The predistortion technology weakens the nonlinear characteristics of the power amplifier of the radiation source, thereby weakening the performance of identifying individual radiation sources. Summary of the Invention

[0003] The purpose of the present invention is to solve the problem that the existing predistortion technology weakens the nonlinear characteristics of the power amplifier of the radiation source, thereby weakening the performance of identifying individual radiation sources, and to propose a method for identifying individual radiation sources based on class activation maps and SincNet networks.

[0004] The specific process of the method for identifying individual radiation sources based on class activation maps and SincNet networks is as follows:

[0005] Judge whether the radiation source signal to be measured is a high signal-to-noise ratio radiation source signal or a low signal-to-noise ratio radiation source signal. If the radiation source signal to be measured is a high signal-to-noise ratio radiation source signal, then execute steps A to D; if the radiation source signal to be measured is a low signal-to-noise ratio radiation source signal, then execute steps one to three;

[0006] The high signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio greater than 15 dB;

[0007] The low signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio less than or equal to 15 dB;

[0008] The specific process is as follows:

[0009] If the radiation source signal to be measured is a high signal-to-noise ratio signal, the specific steps are as follows:

[0010] Step A: The undistorted different radiation source signals v in (n) with labels are input into the predistorter, and the predistorter outputs the undistorted different radiation source signals v pd (n) with labels;

[0011] The undistorted different radiation source signals v pd (n) with labels output by the predistorter are input into the power amplifier, and the power amplifier outputs the signal v pa (n) with labels;

[0012] Step B: Perform CWD processing on the tagged signal v pa (n) to extract the tagged CWD time-frequency distribution feature map;

[0013] Step C: Input the tagged CWD time-frequency distribution feature map extracted in Step B into the ResNet50 network for training to obtain a trained ResNet50 network;

[0014] Step D: Input the untrained distorted radiation source signal to be tested into the predistorter, and the predistorter outputs a predistorted radiation source signal; the predistorted radiation source signal output by the predistorter is input into the power amplifier, and the power amplifier outputs a signal; perform CWD processing on the signal output by the power amplifier to extract the CWD time-frequency distribution feature map;

[0015] Input the extracted CWD time-frequency distribution feature map into the trained ResNet50 network for class activation map analysis to obtain a class activation map function;

[0016] The class activation map function is superimposed and displayed with the CWD time-frequency analysis feature map in the form of a heat map, and the CWD feature map of the sensitive area is analyzed and intercepted;

[0017] Input the CWD feature map of the sensitive area into the support vector machine for recognition;

[0018] If the radiation source signal to be tested is a low signal-to-noise ratio signal, the specific steps are as follows:

[0019] Step 1: Input the tagged undistorted different radiation source signals v in (n) into the predistorter, and the predistorter outputs the tagged predistorted different radiation source signals v pd (n);

[0020] The tagged predistorted different radiation source signals v pd (n) output by the predistorter are input into the power amplifier, and the power amplifier outputs the tagged signal v pa (n);

[0021] The tagged signal v pd (n) output by the predistorter is input into the delay device;

[0022] The tagged signal v pa (n) output by the power amplifier is input into The output signal is input into the digital predistortion trainer, and the digital predistortion trainer outputs G is the power amplifier gain;

[0023] Input the output signal of the digital predistortion trainer and the output signal v of the delay element pd (n) to perform a subtraction operation to output an error e(n). The error e(n) trains the parameters of the digital predistortion trainer model through an adaptive algorithm. When the digital predistortion trainer model is exactly the inverse model of the power amplifier, there is Then the training ends to obtain the corresponding trained digital predistortion trainer; otherwise, the error e(n) continues to train the parameters of the digital predistortion trainer model through the adaptive algorithm until

[0024] Step 2: Input the undistorted different radiation source signals with labels into the predistorter, and the predistorter outputs the predistorted different radiation source signals with labels;

[0025] The predistorted different radiation source signals with labels output by the predistorter are input into the power amplifier, and the power amplifier outputs the signals with labels;

[0026] The signal v pa (n) output by the power amplifier is input The output signal is input into the trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; G is the power amplifier gain;

[0027] The output signal of the trained digital predistortion trainer is input into the SincNet network for training to obtain the trained SincNet network;

[0028] Step 3: Input the undistorted radiation source signal to be tested into the predistorter, and the predistorter outputs the predistorted radiation source signal; the predistorted radiation source signal output by the predistorter is input into the power amplifier, and the power amplifier outputs the signal to be tested;

[0029] The signal to be tested output by the power amplifier is input The output signal is input into the trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; G is the power amplifier gain;

[0030] The output signal of the trained digital predistortion trainer is input into the trained SincNet network to complete the identification of the post-predistortion radiation source individual signal.

[0031] The beneficial effects of the present invention are:

[0032] The present invention relates to a method for analyzing features based on a residual network and applying a class activation map, and a technique for identifying radiation source individuals based on a SincNet network. The former extracts the sensitive region of the time-frequency analysis feature map using sensitivity analysis and uses a support vector machine for identification. At high signal-to-noise ratios, the sensitive region intercepted manually can be directly used for identification, and the same identification accuracy can be obtained without interception, greatly reducing the computational amount and further improving the radiation source individual identification speed. The latter uses SincNet to further accurately extract sensitive frequencies, with a 2.5% increase in the identification rate at low signal-to-noise ratios, and the computational amount is reduced by half compared to ResNext. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the present invention;

[0034] Figure 2 is a diagram of the SincNet network model;

[0035] Figure 3a is the input network feature image of SE1_CWD;

[0036] Figure 3b is the input network feature image of SE2_CWD;

[0037] Figure 3c is the class activation map of SE1;

[0038] Figure 3d is the class activation map of SE2;

[0039] Figure 4a is the average value map of the G channel of the sensitive region of SE1;

[0040] Figure 4b is the average value map of the G channel of the sensitive region of SE2;

[0041] Figure 4c is the average value map of the G channel of the sensitive region of SE3;

[0042] Figure 4d is the average value map of the B channel of the sensitive region of SE1;

[0043] Figure 4e is the average value map of the B channel of the sensitive region of SE2;

[0044] Figure 4f is the average value map of the B channel of the sensitive region of SE3;

[0045] Figure 5 is a schematic diagram of the predistortion structure of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0046] DETAILED DESCRIPTION OF THE INVENTION 1: The specific process of the method for identifying radiation source individuals based on a class activation map and a SincNet network in this embodiment is as follows:

[0047] Determine whether the radiation source signal to be measured is a high signal-to-noise ratio radiation source signal or a low signal-to-noise ratio radiation source signal. If the radiation source signal to be measured is a high signal-to-noise ratio radiation source signal, then perform steps A to D; if the radiation source signal to be measured is a low signal-to-noise ratio radiation source signal, then perform steps one to three;

[0048] The high signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio greater than 15 dB;

[0049] The low signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio less than or equal to 15 dB;

[0050] The specific process is as follows:

[0051] If the radiation source signal to be measured is a high signal-to-noise ratio signal, the specific steps are as follows:

[0052] Step A: The labeled undistorted different radiation source signals v in (n) (high signal-to-noise ratio radiation source signal) are input into the predistorter, and the predistorter outputs the labeled predistorted different radiation source signals v pd (n);

[0053] The labeled predistorted different radiation source signals v pd (n) output by the predistorter are input into the power amplifier (PA), and the power amplifier outputs the labeled signal v pa (n);

[0054] Step B: Perform CWD (Choi-Williams Distributions, CWD) processing on the labeled signal v pa (n) output by the power amplifier to extract the labeled CWD time-frequency distribution feature map;

[0055] Step C: Input the labeled CWD time-frequency distribution feature map extracted in step B into the ResNet50 network for training to obtain the trained ResNet50 network;

[0056] Step D: Input the undistorted radiation source signal to be tested into the predistorter, and the predistorter outputs the predistorted radiation source signal; the predistorted radiation source signal output by the predistorter is input into the power amplifier, and the power amplifier outputs a signal; perform CWD processing on the signal output by the power amplifier to extract the CWD time-frequency distribution feature map;

[0057] Input the extracted CWD time-frequency distribution feature map into the trained ResNet50 network for Class Activation Mapping (CAM) analysis to obtain the class activation map function;

[0058] The class activation map function is superimposed and displayed on the CWD time-frequency analysis feature map in the form of a heat map, and the CWD feature map of the high-sensitivity region is analyzed and intercepted (the high-sensitivity region is manually intercepted by observing the heat map and enclosing the high-sensitivity region with a rectangular box);

[0059] The CWD feature map of the high-sensitivity region is input into a support vector machine for recognition; it is mainly applied to the scenario of individual radiation source recognition under low computational load and high signal-to-noise ratio;

[0060] If the signal of the radiation source to be measured is a low signal-to-noise ratio signal, the specific steps are as follows:

[0061] Step 1: The tagged undistorted different radiation source signals v in (n) (low signal-to-noise ratio radiation source signal) is input into the predistorter, and the predistorter outputs the tagged predistorted different radiation source signals v pd (n);

[0062] The tagged predistorted different radiation source signals v output by the predistorter pd (n) is input into the power amplifier (PA), and the power amplifier outputs the tagged signal v pa (n);

[0063] The tagged signal v output by the predistorter pd (n) is input into the delay device (the delay device is for the purpose of being simultaneous with the output of the digital predistortion trainer); simultaneously;

[0064] The tagged signal v output by the power amplifier pa (n) is input The output signal is input into the digital predistortion trainer, and the digital predistortion trainer outputs G is the power amplifier gain;

[0065] The output signal of the digital predistortion trainer and the output signal v of the delay device pd (n) are subtracted to output the error e(n). The error e(n) trains the model parameters of the digital predistortion trainer through an adaptive algorithm. When the digital predistortion trainer model is exactly the inverse model of the power amplifier, there is Then the training ends to obtain the corresponding trained digital predistortion trainer; otherwise, the error e(n) continues to train the model parameters of the digital predistortion trainer through the adaptive algorithm until

[0066] The predistortion structure adopted is as Figure 5 shown;

[0067] Step 2: Input the untrained and undistorted different radiation source signals with tags into the predistorter, and the predistorter outputs the predistorted different radiation source signals with tags;

[0068] The predistorted different radiation source signals with tags output by the predistorter are input into a power amplifier (PA), and the power amplifier outputs signals with tags;

[0069] The signal v pa (n) output by the power amplifier is input The output signal is input into the trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; G is the power amplifier gain;

[0070] The signal output by the trained digital predistortion trainer is input into the SincNet network for training to obtain a trained SincNet network;

[0071] Step 3: Input the untrained radiation source signal to be tested into the predistorter, and the predistorter outputs the predistorted radiation source signal; the predistorted radiation source signal output by the predistorter is input into the power amplifier, and the power amplifier outputs the signal to be tested;

[0072] The signal to be tested output by the power amplifier is input The output signal is input into the trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; G is the power amplifier gain;

[0073] The signal output by the trained digital predistortion trainer is input into the trained SincNet network to complete the identification of the individual radiation source signals after predistortion.

[0074] The identified categories are different radiation sources. For example, there are several radars, and there are differences between the radars. When a radar emits a radiation signal, the signal is identified to determine which radar the signal belongs to.

[0075] Specific Embodiment 2: The difference between this embodiment and Specific Embodiment 1 is that in step A, the untrained and undistorted different radiation source signals v in (n) are input into the predistorter, and the predistorter outputs the predistorted different radiation source signals v pd (n); the specific process is as follows:

[0076] Based on the QRD-LS algorithm, linearization processing is performed on the power amplifier to obtain the radiation source predistortion signal.

[0077] Other steps and parameters are the same as those in Specific Embodiment 1.

[0078] Embodiment 3: The difference between this embodiment and Embodiment 1 or 2 is that in step A, the tagged pre-distorted different radiation source signal v pd (n) is input into the power amplifier (PA), and the power amplifier outputs the tagged signal v pa (n); The specific expression is:

[0079]

[0080] wherein, v pd (n) is the output signal of the pre-distorter, v pa (n) is the output signal of the power amplifier, n is the signal index, K is the order of the MP model; Q is the depth of the MP model; h kq is the memory polynomial coefficient.

[0081] The MP model is the Memory Polynomial (MP model).

[0082] Other steps and parameters are the same as those in Embodiment 1 or 2.

[0083] Embodiment 4: The difference between this embodiment and any one of Embodiments 1 to 3 is that in step B, the tagged signal v pa (n) output by the power amplifier is subjected to CWD (Choi-Williams Distributions, CWD) processing to extract the tagged CWD time-frequency distribution feature map; The specific process is:

[0084] The CWD distribution is defined as

[0085]

[0086] wherein, CWD(t,f) is the output time-frequency distribution feature map, t is time, f is frequency, σ is the scale factor, τ is the time shift parameter, v is the power amplifier output signal v pa (n), v * is the convolution of the power amplifier output signal v pa (n), and j represents the imaginary number.

[0087] Other steps and parameters are the same as those in any one of Embodiments 1 to 3.

[0088] Embodiment 5: The difference between this embodiment and any one of Embodiments 1 to 4 is that in step D, the to-be-tested non-pre-distorted radiation source signal is input into the pre-distorter, and the pre-distorter outputs the pre-distorted radiation source signal; the pre-distorted radiation source signal output by the pre-distorter is input into the power amplifier, and the power amplifier outputs a signal; the signal output by the power amplifier is subjected to CWD processing to extract the CWD time-frequency distribution feature map;

[0089] The extracted CWD time-frequency distribution feature map is input into the trained ResNet50 network for Class Activation Mapping (CAM) analysis to obtain the class activation map function;

[0090] The class activation map function is superimposed and displayed with the CWD time-frequency analysis feature map in the form of a heat map, and the CWD feature map of the high-sensitivity region (the high-sensitivity region is manually intercepted by observing the heat map and enclosing the high-sensitivity region within a rectangular box) is analyzed and intercepted;

[0091] The CWD feature map of the high-sensitivity region is input into a support vector machine for recognition; it is mainly applied to the scenario of radiation source individual recognition under low computational load and high signal-to-noise ratio;

[0092] The specific process is as follows:

[0093] The to-be-tested undistorted radiation source signal is input into the predistorter, and the predistorter outputs the predistorted radiation source signal; the predistorted radiation source signal output by the predistorter is input into the power amplifier, and the power amplifier outputs a signal; the signal output by the power amplifier is subjected to CWD processing to extract the CWD time-frequency distribution feature map;

[0094] The extracted CWD time-frequency distribution feature map is input into the trained ResNet50 network;

[0095] Let the k feature maps generated by the last convolutional layer in the ResNet50 network be

[0096] where W and b are the weights and learning biases of the neurons in the classification layer; f act is the non-linear activation function of the network; is the tensor product, and x c is the extracted CWD time-frequency distribution feature map;

[0097] The feature map A k is output as

[0098] where is the learning weight corresponding to class c of the output layer unit of the ResNet50 network; y c is the score value with the output class being class c;

[0099] Let the activation value of the k-th unit in the (i, j) position of the last convolutional layer be Accumulate the calculation result

[0100] Define the activation value of the k-th unit at the (i, j) position in the last convolutional layer of the ResNet50 network The weight corresponding to the target class c is

[0101] where Z represents the total number of pixels;

[0102] After the last convolutional layer of the ResNet50 network, there is an output layer, and after the output layer, there is a ReLU activation function;

[0103] After the output layer of the ResNet50 network, a ReLU activation function is set, and the sample x is obtained through the ReLU activation function c Through the class activation map function of the ResNet50 network, the class activation map function is

[0104] The class activation map function is superimposed and displayed with the CWD time-frequency analysis feature map in the form of a heat map, and the CWD feature map of the high-sensitivity area is analyzed and intercepted;

[0105] The CWD feature map of the high-sensitivity area is input into a support vector machine for recognition; it is mainly applied to the scenario of radiation source individual recognition under low computational load and high signal-to-noise ratio.

[0106] Other steps and parameters are the same as those in any one of the first to fifth specific embodiments.

[0107] Specific embodiment six: The difference between this embodiment and the first specific embodiment is that in step one, the labeled undistorted different radiation source signals v in (n) are input into the predistorter, and the predistorter outputs the labeled predistorted different radiation source signals v pd (n); the specific process is:

[0108] Based on the QRD-LS algorithm, linearize the power amplifier to obtain the radiation source predistorted signal.

[0109] Other steps and parameters are the same as those in the first specific embodiment.

[0110] Specific embodiment seven: The difference between this embodiment and the sixth specific embodiment is that in step one, the labeled predistorted different radiation source signals v pd (n) output by the predistorter are input into the power amplifier (PA), and the power amplifier outputs the labeled signal v pa (n); the specific expression is:

[0111]

[0112] where v pd (n) is the output signal of the predistorter, v pa(n) is the output signal of the power amplifier, n is the signal index, K is the order of the MP model; Q is the depth of the MP model; h kq is the memory polynomial coefficient;

[0113] The MP model is a memory polynomial model (Memory Polynomial, MP model);

[0114] The tagged signal v output by the power amplifier in step one pa (n) is input The output signal is input to the digital predistortion trainer, and the digital predistortion trainer outputs G is the power amplifier gain; the specific expression is:

[0115]

[0116] Among them, is the output signal of the digital predistortion trainer, v pa is the output signal of the power amplifier, n is the signal index, K is the order of the MP model; Q is the depth of the MP model; G is the power amplifier gain; w kq is the digital predistortion coefficient.

[0117] The digital predistortion coefficient w kq is the solution of.

[0118] Other steps and parameters are the same as those in the sixth specific implementation manner.

[0119] Specific implementation manner eight: The difference between this implementation manner and the seventh specific implementation manner is that the SincNet network in step five sequentially includes: an input layer, a filter structure, a first pooling layer, a first convolutional layer, a second pooling layer, a second convolutional layer, a third pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, a fourth fully connected layer, a LeakyReLU activation function, and a softmax layer.

[0120] Other steps and parameters are the same as those in the seventh specific implementation manner.

[0121] Specific implementation manner nine: The difference between this implementation manner and the eighth specific implementation manner is that the filter structure is an 80-dimensional filter structure, the starting bandwidth of each filter is about 0.37 MHz, and the starting frequency and cut-off frequency of the filter structure are continuously distributed from 0 to 30 MHz.

[0122] Other steps and parameters are the same as those in the eighth specific implementation manner.

[0123] Specific implementation manner ten: The difference between this implementation manner and the ninth specific implementation manner is that the first pooling layer is a pooling layer with a window size of 3;

[0124] The convolution kernels of the first convolutional layer and the second convolutional layer are both 5×5;

[0125] The second pooling layer and the third pooling layer are pooling layers with a window size of 3;

[0126] After each of the first fully-connected layer, the second fully-connected layer, the third fully-connected layer, and the fourth fully-connected layer, a BN layer is connected.

[0127] Other steps and parameters are the same as those in the ninth specific implementation manner.

[0128] As Figure 2 , the SincNet filter structure is essentially to filter and extract more effective low-dimensional features in the first layer of the network to achieve the extraction of frequency-sensitive regions. Its behavior of performing sensitivity analysis with the class activation map and selecting the sensitive region has the same purpose. At a sampling rate of 60 MHz, the first layer SincNet of the network is a filter structure with a total of 80 dimensions. The starting bandwidth of each filter is about 0.37 MHz, and the starting frequency and cut-off frequency of the structure are continuously distributed from 0 to 30 MHz. On this basis, a pooling layer with a window size of 3 is used for dimensionality reduction. The output passes through two pairs of convolutional layers with a convolution kernel size of 5×5 and a pooling layer with a window size of 3 to embed the underlying features of the samples, and then through four fully-connected layers. A BN layer is also set after the fully-connected layer, the LeakyReLU activation function is used, the Dropout ratio is set to 0.5, and finally, after normalization by the softmax layer, the individual signal recognition of the post-distortion radiation source is completed. It is mainly applied to the scenario of individual radiation source recognition with more stringent signal-to-noise ratio.

[0129] The following embodiments are used to verify the beneficial effects of the present invention:

[0130] Embodiment 1:

[0131] The feature images obtained by performing CWD transformation on two post-distortion radiation source individuals are input into the obtained ResNet50-CWD network. The weights for classifying the feature maps are backpropagated to the input time-frequency spectrum through the activation function, and finally, the class activation maps of the trained network are obtained as shown in Figure 3a 、 3b 、3c, 3d, Figure 3a 、 3b, in 3c and 3d, the bright areas represent the parts with high classification weights in that area, which play an important role in network recognition. It can be found that the bright areas are mainly distributed along the time-frequency part of the main signal in the time-frequency diagram, while the dark areas are mainly distributed in the noise parts of non-main frequencies. It can be seen that the analysis of sensitive areas of the Grad-CAM method for time-frequency diagrams has practical physical significance. By observing and analyzing the distribution of sensitive areas for 500 samples of each of the three radiation source individuals, it is found that the sensitive areas of different samples of the same individual converge, while the distribution positions of the sensitive areas of the three radiation sources are significantly different. Taking SE1 and SE2 as examples, the main sensitive areas of the class activation map of SE1 are distributed in the edge part of the modulation, while SE2 is mainly distributed in the intermediate frequency part.

[0132] Perform a secondary analysis on the characteristics of the local information of the sensitive areas. For the sensitive areas of the class activation map of SE1, intercept the corresponding areas of 1500 feature map samples, and explore the class separability of the local features of the radiation source signal at the sensitive areas after pre-distortion, as Figure 4a , 4b shown in 4c, 4d, 4e, and 4f. After summing up each channel of the sensitive areas of 1500 samples of the three radiation sources and taking the average, the time-frequency images of the GB two channels are drawn. It can be found that for different radiation sources, the differences in sensitive areas are obvious, especially in the B channel. Generally, the between-class scatter is strong.

[0133] Directly use the sensitive areas of the feature maps for high-precision recognition. Use SVM as the classifier, with SE1, SE2, and SE3 as the recognition objects, and the number of samples after pre-distortion is 4500. Compare the recognition accuracy of traditional features under SVM, as shown in Table 1. The results show that using high-sensitivity area recognition can obtain a recognition accuracy comparable to that of using all areas at high signal-to-noise ratios, but at low signal-to-noise ratios, the recognition accuracy will decrease sharply, with a 9.2% decrease in recognition accuracy at -5 dB. This result is reasonable. Because the high-sensitivity areas in this part are rectangular windows manually intercepted after observation and analysis, which are not completely correct. In fact, the real sensitive areas should be irregular and discrete areas. On the other hand, using high-sensitivity area recognition will inevitably reduce the computational amount, which is a major advantage of this method.

[0134] Table 1 Recognition accuracy of SVM using different features

[0135]

[0136] Example 2:

[0137] Use a neural network based on the SincNet structure to identify the signal time series after pre-distorting the input numbers. Use 500 samples each of SE1, SE2, and SE3, and train with 1000 sampling points of single pulses under a total of 1500 samples. Use 1000 sampling points of pulses in different time periods of the same 1500 samples for testing. Set an adaptive learning rate function, with a batch input of 128 samples, each sample having 1000 sampling points. Finally, the filter frequency band learned by the first layer of SincNet is the signal sensitive frequency. Verify the performance of the network constructed in this paper. The final recognition accuracy at different signal-to-noise ratios and the results compared with other networks are shown in Table 2. It is found that compared with ResNeXt50-CWD, the most accurate method among the comparison methods, the SincNet network with pre-distortion has a great advantage in the individual recognition of radiation sources under strong noise. The accuracy is improved by 2.5% at -5dB, and the recognition accuracies of the two are comparable under high signal-to-noise ratios. Because the SincNet filter structure realizes the accurate selection of sensitive regions and can better filter out noise interference, it has better results under low signal-to-noise ratios. Compare the total floating-point operations of the algorithm and the total number of algorithm model parameters, and the results are shown in Table 3. It can be seen from the table that the ResNeXt50-CWD method has a large amount of computation, which is caused by its complex network structure. The method based on the SincNet network has a reduction of half in the amount of computation and parameters, because the network directly filters in the first layer, learns filter parameters, pays more attention to sensitive regions, and reduces the network complexity.

[0138] Table 2 Recognition Accuracy of Different Networks at Different Signal-to-Noise Ratios

[0139]

[0140] Table 3 Comparison of Computation Amounts

[0141]

[0142] The present invention can also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and deformations according to the present invention, but these corresponding changes and deformations should all fall within the protection scope of the appended claims of the present invention.

Claims

1. A method for identifying individual radiation sources based on class activation maps and SincNet networks, characterized in that: The specific process of the method is as follows: Determine whether the radiation source signal to be measured is a high signal-to-noise ratio radiation source signal or a low signal-to-noise ratio radiation source signal. If the radiation source signal to be measured is a high signal-to-noise ratio radiation source signal, then execute steps A to D; if the radiation source signal to be measured is a low signal-to-noise ratio radiation source signal, then execute steps one to three. The high signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio greater than 15 db. The low signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio less than or equal to 15 db. The specific process is as follows: If the radiation source signal to be measured is a high signal-to-noise ratio signal, the specific steps are as follows: Step A, tagged non-pre-distorted different radiation source signals v in (n) enter a pre-distorter, and the output of the pre-distorter is tagged pre-distorted different radiation source signals v pd (n); Tagged pre-distorted different radiation source signals v output by the pre-distorter pd (n) Input power amplifier, and the power amplifier outputs tagged signal v pa (n); Step B: Perform CWD processing on the tagged signal v pa (n) output by the power amplifier to extract the tagged CWD time-frequency distribution feature map; Step C: Input the labeled CWD time-frequency distribution feature map extracted in step B into the ResNet50 network for training to obtain a trained ResNet50 network. Step D: Input the undistorted radiation source signal to be tested into the predistorter, and the predistorter outputs a predistorted radiation source signal; the predistorted radiation source signal output by the predistorter is input into the power amplifier, and the power amplifier outputs a signal; perform CWD processing on the signal output by the power amplifier to extract the CWD time-frequency distribution feature map. Input the extracted CWD time-frequency distribution feature map into the trained ResNet50 network for class activation map analysis to obtain a class activation map function. The class activation map function is superimposed and displayed with the CWD time-frequency analysis feature map in the form of a heat map, and analyze and intercept the CWD feature map of the sensitive area. Input the CWD feature map of the sensitive area into the support vector machine for recognition. If the radiation source signal to be measured is a low signal-to-noise ratio signal, the specific steps are as follows: Step 1: Tagged non-pre-distorted different radiation source signals v in (n) Enter the pre-distorter, and the output of the pre-distorter is the tagged pre-distorted different radiation source signals v pd (n); Tagged pre-distorted different radiation source signals v output by the pre-distorter pd (n) Input power amplifier, and the power amplifier outputs tagged signal v pa (n); The tagged signal v output by the predistorter pd (n) input delay element; The labeled signal v output by the power amplifier pa (n) Input The output signal is input to the digital predistortion trainer, and the digital predistortion trainer outputs G is the power amplifier gain; The output signal of the digital predistortion trainer and the output signal v pd (n) are subtracted to output the error e(n). The error e(n) trains the model parameters of the digital predistortion trainer through an adaptive algorithm. When the digital predistortion trainer model is exactly the inverse model of the power amplifier, there is Then the training ends to obtain the corresponding trained digital predistortion trainer; otherwise, the error e(n) continues to train the model parameters of the digital predistortion trainer through the adaptive algorithm until Step two: Input the labeled undistorted different radiation source signals into the predistorter, and the predistorter outputs labeled predistorted different radiation source signals. The labeled predistorted different radiation source signals output by the predistorter are input into the power amplifier, and the power amplifier outputs a labeled signal. Tagged signal v output by the power amplifier pa (n) Input The output signal is input to the trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; G is the power amplifier gain; Input the output signal of the trained digital predistortion trainer into the SincNet network for training to obtain a trained SincNet network. Step three: Input the undistorted radiation source signal to be tested into the predistorter, and the predistorter outputs a predistorted radiation source signal; the predistorted radiation source signal output by the predistorter is input into the power amplifier, and the power amplifier outputs the signal to be tested. Input of the signal to be tested output by the power amplifier The output signal is input into the trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; G is the power amplifier gain; Input the output signal of the trained digital predistortion trainer into the trained SincNet network to complete the recognition of the individual radiation source signals after predistortion.

2. The method for identifying individual radiation sources based on class activation maps and SincNet networks according to claim 1, wherein: The untwisted different radiation source signals v with tags in step A in (n0 are input into a predistorter, and the predistorter outputs the twisted different radiation source signals v pd (n0; the specific process is as follows: Based on the QRD-LS algorithm, linearize the power amplifier to obtain the radiation source predistorted signal.

3. The method for identifying radiation source individuals based on class activation maps and SincNet networks according to claim 2, wherein: The tagged pre-distorted different radiation source signals v output by the pre-distorter in the step A pd (n) are input to a power amplifier, and the power amplifier outputs a tagged signal v pa (n); the specific expression is: where, v pd (n) is the output signal of the predistorter, v pa (n) is the output signal of the power amplifier, n is the signal index, K is the order of the MP model; Q is the depth of the MP model; h kq is the memory polynomial coefficient; The MP model is a memory polynomial model.

4. The method for identifying individual radiation sources based on class activation maps and SincNet networks according to claim 3, characterized in that: In step B, the tagged signal v pa (n) output by the power amplifier is subjected to CWD processing to extract the tagged CWD time-frequency distribution feature map. The specific process is as follows: The CWD distribution is defined as Among them, CWD(t, f) is the output time-frequency distribution feature map, where t is time, f is frequency, σ is the scale factor, τ is the time shift parameter, and v is the output signal v of the power amplifier pa (n), v * is the convolution of the output signal v of the power amplifier pa (n0, and j represents the imaginary number.

5. The method for identifying radiation source individuals based on class activation maps and SincNet networks according to claim 4, wherein: In step D, input the undistorted radiation source signal to be tested into the predistorter, and the predistorter outputs a predistorted radiation source signal; the predistorted radiation source signal output by the predistorter is input into the power amplifier, and the power amplifier outputs a signal; perform CWD processing on the signal output by the power amplifier to extract the CWD time-frequency distribution feature map. Input the extracted CWD time-frequency distribution feature map into the trained ResNet50 network for class activation map analysis to obtain a class activation map function. The class activation map function is superimposed on the CWD time-frequency analysis feature map in the form of a heat map to analyze and intercept the CWD feature map of the sensitive region; The CWD feature map of the high-sensitivity region is input into a support vector machine for recognition; The specific process is as follows: The undistorted radiation source signal to be tested is input into the predistorter, and the predistorter outputs the predistorted radiation source signal; the predistorted radiation source signal output by the predistorter is input into the power amplifier, and the power amplifier outputs a signal; the signal output by the power amplifier is subjected to CWD processing to extract the CWD time-frequency distribution feature map; The extracted CWD time-frequency distribution feature map is input into the trained ResNet50 network; Let the k feature maps generated by the last convolutional layer in the ResNet50 network be Among them, W and b are the weights and learning biases of the neurons in the classification layer; f act is the non-linear activation function of the network; is the tensor product, and x c is the extracted CWD time-frequency distribution feature map; Feature map A k Output after passing through the output layer of the ResNet50 network is Among them, is the learning weight corresponding to class c of the output layer unit of the ResNet50 network; y c is the score value with the output category being class c. Define the activation value of the k-th unit in the last convolutional layer of the ResNet50 network at the position (i, j0) The weight corresponding to the target class c is Among them, Z represents the total number of pixels; Set a ReLU activation function after the output layer of the ResNet50 network, and obtain the sample x through the ReLU activation function c After passing through the class activation map function of the ResNet50 network, the class activation map function is The class activation map function is superimposed on the CWD time-frequency analysis feature map in the form of a heat map to analyze and intercept the CWD feature map of the sensitive region; The CWD feature map of the sensitive region is input into a support vector machine for recognition.

6. The method for identifying individual radiation sources based on class activation maps and SincNet networks according to claim 5, wherein: The untwisted different radiation source signals v with tags in the first step in (n0 are input into the predistorter, and the predistorter outputs the predistorted different radiation source signals v pd (n0; the specific process is as follows: Based on the QRD-LS algorithm, the power amplifier is linearized to obtain the predistorted signal of the radiation source.

7. The method for identifying individual radiation sources based on class activation maps and SincNet networks according to claim 6, characterized in that: The pre-distorted different radiation source signals v with tags output by the pre-distorter in the first step pd (n) are input to the power amplifier, and the power amplifier outputs the signal v pa (n); the specific expression is: Among them, v pd (n) is the output signal of the predistorter, v pa (n) is the output signal of the power amplifier, n is the signal index, K is the order of the MP model; Q is the depth of the MP model; h kq is the memory polynomial coefficient; The MP model is a memory polynomial model; The tagged signal v pa (n) output by the power amplifier in the first step is input The output signal is input into the digital pre-distortion trainer, and the digital pre-distortion trainer outputs G is the power amplifier gain; the specific expression is: Among them, is the output signal of the digital pre-distortion trainer, v pa is the output signal of the power amplifier, n is the signal index, K is the order of the MP model; Q is the depth of the MP model; G is the power amplifier gain; w kq is the digital pre-distortion coefficient.

8. The method for identifying radiation source individuals based on class activation maps and SincNet networks according to claim 7, wherein: The SincNet network sequentially includes: an input layer, a filter structure, a first pooling layer, a first convolutional layer, a second pooling layer, a second convolutional layer, a third pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, a fourth fully connected layer, a LeakyReLU activation function, and a softmax layer.

9. The method for identifying radiation source individuals based on class activation maps and SincNet networks according to claim 8, wherein: The filter structure is an 80-dimensional filter structure, and the start frequency and cut-off frequency of the filter structure are continuously distributed from 0 to 30 MHz.

10. The method for identifying individual radiation sources based on class activation maps and SincNet networks according to claim 9, wherein: The first pooling layer is a pooling layer with a window size of 3; The convolutional kernel sizes of the first convolutional layer and the second convolutional layer are both 5×5; The second pooling layer and the third pooling layer are pooling layers with a window size of 3; After each of the first fully connected layer, the second fully connected layer, the third fully connected layer, and the fourth fully connected layer in the first fully connected layer, a BN layer is connected.

Images