CN115270878A - Radiation source individual identification method based on class activation diagram and SincNet network - Google Patents

Radiation source individual identification method based on class activation diagram and SincNet network Download PDF

Info

Publication number
CN115270878A
CN115270878A CN202210903562.1A CN202210903562A CN115270878A CN 115270878 A CN115270878 A CN 115270878A CN 202210903562 A CN202210903562 A CN 202210903562A CN 115270878 A CN115270878 A CN 115270878A
Authority
CN
China
Prior art keywords
signal
radiation source
power amplifier
output
cwd
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210903562.1A
Other languages
Chinese (zh)
Inventor
赵雅琴
丁沁宇
吴龙文
何胜阳
韩易伸
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Institute of Technology
Original Assignee
Harbin Institute of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Harbin Institute of Technology filed Critical Harbin Institute of Technology
Priority to CN202210903562.1A priority Critical patent/CN115270878A/en
Publication of CN115270878A publication Critical patent/CN115270878A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
    • G01S7/021—Auxiliary means for detecting or identifying radar signals or the like, e.g. radar jamming signals
    • H—ELECTRICITY
    • H03—ELECTRONIC CIRCUITRY
    • H03F—AMPLIFIERS
    • H03F1/00—Details of amplifiers with only discharge tubes, only semiconductor devices or only unspecified devices as amplifying elements
    • H03F1/32—Modifications of amplifiers to reduce non-linear distortion
    • H03F1/3241—Modifications of amplifiers to reduce non-linear distortion using predistortion circuits
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/309—Measuring or estimating channel quality parameters
    • H04B17/336—Signal-to-interference ratio [SIR] or carrier-to-interference ratio [CIR]

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • General Physics & Mathematics (AREA)
  • Quality & Reliability (AREA)
  • Electromagnetism (AREA)
  • Signal Processing (AREA)
  • Nonlinear Science (AREA)
  • Power Engineering (AREA)
  • Amplifiers (AREA)

Abstract

The invention relates to a radiation source individual identification method based on a class activation diagram and a SincNet network. The invention aims to solve the problems that the existing predistortion technology weakens the nonlinear characteristics of a power amplifier of a radiation source and further weakens the individual identification performance of the radiation source. Judging whether the radiation source signal to be detected is a high signal-to-noise ratio radiation source signal or a low signal-to-noise ratio radiation source signal, and executing A to D if the radiation source signal to be detected is the high signal-to-noise ratio radiation source signal; if the radiation source signal to be detected is a low signal-to-noise ratio radiation source signal, executing one to three; A. the power amplifier outputs a signal with a label; b: extracting a CWD time-frequency distribution characteristic diagram with a label; c: obtaining a trained ResNet50 network; d: completing individual signals of the radiation source after pre-distortion to be detected; firstly, the method comprises the following steps: obtaining a trained digital predistortion trainer; II, secondly: obtaining a trained SincNet network; thirdly, the steps of: and finishing the identification of individual signals of the radiation source after the pre-distortion to be detected. The invention is used in the field of individual identification of radiation sources.

Description

Radiation source individual identification method based on class activation diagram and SincNet network
Technical Field
The invention relates to a radiation source individual identification method.
Background
Radiation source individual Identification (SEI) techniques achieve Identification discrimination of radiation sources primarily based on the unintentional modulation characteristics of the signal. The nonlinear distortion of the power amplifier is one of the unintentional modulation characteristics and one of the main bases for individual identification of the radiation source, but in communication and radar systems, the nonlinear distortion of the power amplifier causes the reduction of the linear range and the efficiency. With the expanding requirements of broadband communication and radar systems, power amplifier linearization technologies such as predistortion and the like appear. The predistortion technology weakens the nonlinear characteristics of the power amplifier of the radiation source, and further weakens the individual identification performance of the radiation source.
Disclosure of Invention
The invention aims to solve the problems that the existing predistortion technology weakens the nonlinear characteristics of a power amplifier of a radiation source and further weakens the individual identification performance of the radiation source, and provides a radiation source individual identification method based on a class activation diagram and a SincNet network.
The radiation source individual identification method based on the class activation graph and the SincNet network comprises the following specific processes:
judging whether the radiation source signal to be detected is a high signal-to-noise ratio radiation source signal or a low signal-to-noise ratio radiation source signal, and executing the step A to the step D if the radiation source signal to be detected is the high signal-to-noise ratio radiation source signal; if the radiation source signal to be detected is a low signal-to-noise ratio radiation source signal, executing the first step to the third step;
the high signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio of more than 15 db;
the low signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio of less than or equal to 15 db;
the specific process is as follows:
if the radiation source signal to be detected is a high signal-to-noise ratio signal, the specific steps are as follows:
step A, carrying out label-carrying non-predistortion different radiation source signals vin(n) input to a predistorter which outputs tagged predistorted different radiation source signals vpd(n)；
Output signal v of predistorter with labeled predistortion different radiation sourcepd(n) input to a power amplifier, which outputs a tagged signal vpa(n)；
And B: tagged signal v to the output of a power amplifierpa(n) performing CWD processing, and extracting a marked CWD time-frequency distribution characteristic diagram;
step C: inputting the marked CWD time-frequency distribution characteristic diagram extracted in the step B into a ResNet50 network for training to obtain a trained ResNet50 network;
step D: inputting an un-predistorted radiation source signal to be tested into a predistorter, and outputting a predistorted radiation source signal by the predistorter; a predistortion radiation source signal output by the predistorter is input into a power amplifier, and the power amplifier outputs a signal; performing CWD processing on the signal output by the power amplifier, and extracting a CWD time-frequency distribution characteristic diagram;
inputting the extracted CWD time-frequency distribution characteristic graph into a trained ResNet50 network, and analyzing a class activation graph to obtain a class activation graph function;
the class activation graph function is displayed in an overlaying mode with the CWD time-frequency analysis feature graph in the form of thermodynamic diagram, and the CWD feature graph of the sensitive area is analyzed and intercepted;
inputting a CWD characteristic diagram of the sensitive area into a support vector machine for identification;
if the radiation source signal to be detected is a low signal-to-noise ratio signal, the specific steps are as follows:
the method comprises the following steps: labeled unpredistorted different radiation source signals vin(n) input to a predistorter which outputs tagged predistorted different radiation source signals vpd(n)；
Tagged predistorter outputDistorting different radiation source signals vpd(n) input to a power amplifier, which outputs a tagged signal vpa(n)；
Tagged signal v output by predistorterpd(n) an input delay;
tagged signal v output by a power amplifierpa(n) input
Figure BDA0003769902760000021
The output signal is input into the digital predistortion trainer, and the output of the digital predistortion trainer
Figure BDA0003769902760000022
G is the power amplifier gain;
output signal of digital predistortion trainer
Figure BDA0003769902760000023
And the delay output signal vpd(n) carrying out subtraction operation to output an error e (n), wherein the error e (n) trains the model parameters of the digital predistortion trainer through a self-adaptive algorithm, and when the digital predistortion trainer model is just the inverse model of the power amplifier, the error e (n) has the following characteristics that
Figure BDA0003769902760000024
Obtaining a corresponding trained digital predistortion trainer after the training is finished; otherwise, the error e (n) continues to train the model parameters of the digital predistortion trainer through the adaptive algorithm until the model parameters of the digital predistortion trainer are trained
Figure BDA0003769902760000025
Step two: inputting the different radiation source signals with labels without predistortion into a predistorter, and outputting the different radiation source signals with labels with predistortion by the predistorter;
the pre-distorter outputs different pre-distorted radiation source signals with labels to the power amplifier, and the power amplifier outputs the signals with labels;
tagged signal v output by a power amplifierpa(n) input
Figure BDA0003769902760000026
The output signal is input into a trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; g is the power amplifier gain;
inputting the output signal of the trained digital predistortion trainer into a SincNet network for training to obtain the trained SincNet network;
step three: inputting an un-predistorted radiation source signal to be tested into a predistorter, and outputting a predistorted radiation source signal by the predistorter; a predistortion radiation source signal output by the predistorter is input into a power amplifier, and the power amplifier outputs a signal to be tested;
signal input to be tested output by power amplifier
Figure BDA0003769902760000031
The output signal is input into a trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; g is the power amplifier gain;
and inputting the output signal of the trained digital predistortion trainer into the trained SincNet network to finish the identification of the individual signal of the radiation source after predistortion.
The invention has the beneficial effects that:
the invention relates to a method for analyzing characteristics by using a class activation graph based on a residual error network and a radiation source individual identification technology based on a SincNet network. The method utilizes sensitivity analysis to extract a time-frequency analysis characteristic diagram sensitivity area and utilizes a support vector machine to identify, can directly utilize the sensitivity area identification of manual interception under high signal-to-noise ratio, and obtains equivalent identification precision when not intercepted, thereby greatly reducing the calculated amount and further improving the individual identification speed of the radiation source. The sensitive frequency is further accurately extracted by using SincNet, the recognition rate is improved by 2.5% under the condition of low signal-to-noise ratio, and the calculated amount is reduced by half compared with ResNext.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a SincNet network model diagram;
FIG. 3a is a SE1_ CWD input network feature image;
FIG. 3b is a SE2_ CWD input network feature image;
FIG. 3c is a SE1 class activation map;
FIG. 3d is a SE2 class activation map;
FIG. 4a is a graph of the average value of the G channel of the SE1 sensitive region;
FIG. 4b is a graph of the average value of the G channel of the SE2 sensitive region;
FIG. 4c is a graph of the average value of the G channel of the SE3 sensitive region;
FIG. 4d is a graph of the B channel mean values of the SE1 sensitive region;
FIG. 4e is a graph of the average of B channels in the SE2 sensitive region;
FIG. 4f is a graph of the B channel mean values of the SE3 sensitive region;
fig. 5 is a diagram illustrating a predistortion structure according to the present invention.
Detailed Description
The first specific implementation way is as follows: the method for identifying the individual radiation source based on the class activation graph and the SincNet network comprises the following steps:
judging whether the radiation source signal to be detected is a high signal-to-noise ratio radiation source signal or a low signal-to-noise ratio radiation source signal, and executing the step A to the step D if the radiation source signal to be detected is the high signal-to-noise ratio radiation source signal; if the radiation source signal to be detected is a low signal-to-noise ratio radiation source signal, executing the first step to the third step;
the high signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio of more than 15 db;
the low signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio of less than or equal to 15 db;
the specific process is as follows:
if the radiation source signal to be detected is a high signal-to-noise ratio signal, the specific steps are as follows:
step A, carrying out label-carrying non-predistortion different radiation source signals vin(n) (high SNR radiation source signal) is input into a predistorter, which outputs a labeled predistorted different radiation source signal vpd(n)；
Tagged predistorted differential radiation output by predistorterSource signal vpd(n) input to a Power Amplifier (PA) which outputs a tagged signal vpa(n)；
And B, step B: for the tagged signal v output by the power amplifierpa(n) performing CWD (Choi-Williams Distributions, CWD) processing, and extracting a marked CWD time-frequency distribution characteristic diagram;
and C: inputting the marked CWD time-frequency distribution characteristic diagram extracted in the step B into a ResNet50 network for training to obtain a trained ResNet50 network;
step D: inputting an un-predistorted radiation source signal to be tested into a predistorter, and outputting a predistorted radiation source signal by the predistorter; a predistortion radiation source signal output by the predistorter is input into a power amplifier, and the power amplifier outputs a signal; performing CWD processing on the signal output by the power amplifier, and extracting a CWD time-frequency distribution characteristic diagram;
inputting the extracted CWD time-frequency distribution characteristic diagram into a trained ResNet50 network, and performing Class Activation Mapping (CAM) analysis to obtain a Class Activation diagram function;
the class activation graph function is displayed in an overlaying mode with the CWD time-frequency analysis feature graph in a thermodynamic diagram mode, and the CWD feature graph of a high-sensitivity area (the high-sensitivity area is selected by manually intercepting a rectangular frame through observing the thermodynamic diagram) is analyzed and intercepted;
inputting the CWD characteristic diagram of the high-sensitivity region into a support vector machine for identification; the method is mainly applied to individual identification scenes of the radiation source under low computation quantity and high signal-to-noise ratio;
if the radiation source signal to be detected is a low signal-to-noise ratio signal, the specific steps are as follows:
the method comprises the following steps: labeled non-pre-distorted different radiation source signals vin(n) (low signal-to-noise ratio radiation source signal) is input to a predistorter, which outputs a tagged, predistorted different radiation source signal vpd(n)；
Tagged predistortion diverse radiation source signals v output by predistorterpd(n) input to a Power Amplifier (PA) which outputs a tagged signal vpa(n)；
Tagged signal v output by predistorterpd(n) input delay (delay for output from digital predistortion trainer)
Figure BDA0003769902760000051
At the same time);
tagged signal v output by a power amplifierpa(n) input
Figure BDA0003769902760000052
The output signal is input into the digital predistortion trainer, and the output of the digital predistortion trainer
Figure BDA0003769902760000053
G is the power amplifier gain;
output signal of digital predistortion trainer
Figure BDA0003769902760000054
And the delay output signal vpd(n) carrying out subtraction operation to output an error e (n), wherein the error e (n) trains the model parameters of the digital predistortion trainer through a self-adaptive algorithm, and when the digital predistortion trainer model is just the inverse model of the power amplifier, the error e (n) has the following characteristics that
Figure BDA0003769902760000055
Obtaining a corresponding trained digital predistortion trainer after the training is finished; otherwise, the error e (n) continues to train the model parameters of the digital predistortion trainer through the adaptive algorithm until the model parameters of the digital predistortion trainer are trained
Figure BDA0003769902760000056
The predistortion structure used is shown in fig. 5;
step two: inputting the different radiation source signals with labels without predistortion into a predistorter, and outputting the different radiation source signals with labels with predistortion by the predistorter;
the pre-distorted different radiation source signals with labels output by the pre-distorter are input into a Power Amplifier (PA), and the power amplifier outputs signals with labels;
tagged signal v output by a power amplifierpa(n) input
Figure BDA0003769902760000057
The output signal is input into a trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; g is the power amplifier gain;
inputting the output signal of the trained digital predistortion trainer into a SincNet network for training to obtain a trained SincNet network;
step three: inputting an un-predistorted radiation source signal to be tested into a predistorter, and outputting a predistorted radiation source signal by the predistorter; a predistortion radiation source signal output by the predistorter is input into a power amplifier, and the power amplifier outputs a signal to be tested;
signal input to be tested output by power amplifier
Figure BDA0003769902760000058
The output signal is input into a trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; g is the power amplifier gain;
and inputting the output signal of the trained digital predistortion trainer into the trained SincNet network to finish the identification of the individual signal of the radiation source after predistortion.
The identified categories are different radiation sources. There are, for example, several radars, with a difference between the radars. When the radar radiates a signal, the signal is identified, and the radar to which the signal belongs is judged.
The second embodiment is as follows: in a difference from the first embodiment, the labeled non-pre-distorted different radiation source signal v in step ain(n) input to a predistorter which outputs tagged predistorted different radiation source signals vpd(n); the specific process is as follows:
and (3) carrying out linearization treatment on the power amplifier based on a QRD-LS algorithm to obtain a radiation source predistortion signal.
Other steps and parameters are the same as those in the first embodiment.
The third concrete implementation mode: in this embodiment, different from the first or second embodiment, the pre-distorter outputs the pre-distorter in step a, the pre-distorts the different radiation source signal v with the tagpd(n) input to a Power Amplifier (PA) which outputs a tagged signal vpa(n); the specific expression is as follows:
Figure BDA0003769902760000061
wherein v ispd(n) is the predistorter output signal, vpa(n) is the output signal of the power amplifier, n is the signal index, and K is the MP model order; q is the MP model depth; h is a total ofkqIs a memory polynomial coefficient.
The MP model is a Memory Polynomial model (MP model).
Other steps and parameters are the same as those in the first or second embodiment.
The fourth concrete implementation mode is as follows: this embodiment is different from the first to third embodiments in that the labeled signal v outputted to the power amplifier in the step Bpa(n) performing CWD (Choi-Williams Distributions, CWD) processing, and extracting a marked CWD time-frequency distribution characteristic diagram; the specific process is as follows:
the CWD distribution is defined as
Figure BDA0003769902760000062
Wherein, CWD (t, f) is the output time-frequency distribution characteristic diagram, t is time, f is frequency, sigma is scale factor, tau is time shift parameter, v is the output signal v of the power amplifierpa(n)，v*For outputting a signal v to a power amplifierpa(n) and j represents an imaginary number.
Other steps and parameters are the same as those in one of the first to third embodiments.
The fifth concrete implementation mode is as follows: the difference between this embodiment and one of the first to the fourth embodiments is that, in step D, the non-predistortion radiation source signal to be tested is input into the predistorter, and the predistorter outputs a predistortion radiation source signal; a predistortion radiation source signal output by the predistorter is input into a power amplifier, and the power amplifier outputs a signal; performing CWD processing on the signal output by the power amplifier, and extracting a CWD time-frequency distribution characteristic diagram;
inputting the extracted CWD time-frequency distribution characteristic diagram into a trained ResNet50 network, and performing Class Activation Mapping (CAM) analysis to obtain a Class Activation diagram function;
the class activation graph function is displayed in an overlaying mode with the CWD time-frequency analysis feature graph in a thermodynamic diagram mode, and the CWD feature graph of a high-sensitivity area (the high-sensitivity area is selected by manually intercepting a rectangular frame through observing the thermodynamic diagram) is analyzed and intercepted;
inputting the CWD characteristic diagram of the high sensitivity area into a support vector machine for identification; the method is mainly applied to individual identification scenes of the radiation source under low computation quantity and high signal-to-noise ratio;
the specific process is as follows:
inputting an un-predistorted radiation source signal to be tested into a predistorter, and outputting a predistorted radiation source signal by the predistorter; a predistortion radiation source signal output by the predistorter is input into a power amplifier, and the power amplifier outputs a signal; performing CWD processing on the signal output by the power amplifier, and extracting a CWD time-frequency distribution characteristic diagram;
inputting the extracted CWD time-frequency distribution characteristic diagram into a trained ResNet50 network;
let k feature maps generated by the last convolutional layer in the ResNet50 network be
Figure BDA0003769902760000071
Wherein W and b are the weight and learning bias of the classification layer neurons, respectively; f. ofactA nonlinear activation function for the network;
Figure BDA0003769902760000072
is the product of the tensors, xcThe extracted CWD time frequency distribution characteristic graph is obtained;
characteristic diagram AkOutput through ResNet50 network output layer as
Figure BDA0003769902760000073
Wherein the content of the first and second substances,
Figure BDA0003769902760000074
c-class learning weights corresponding to the ResNet50 network output layer units; y iscA score value with an output category of c is obtained;
setting the activation value of the kth cell of the last convolutional layer at the (i, j) position to be
Figure BDA0003769902760000075
Accumulating the calculation results
Figure BDA0003769902760000076
Defining the activation value of the kth cell of the last convolutional layer in the ResNet50 network at position (i, j)
Figure BDA0003769902760000077
The weight corresponding to the object class c is
Figure BDA0003769902760000078
Wherein Z represents the total number of pixels;
the ResNet50 network is followed by an output layer after the last convolutional layer, and the output layer is followed by a ReLU activation function;
setting a ReLU activation function after the ResNet50 network output layer, and obtaining a sample x through the ReLU activation functioncClass activation graph function through ResNet50 network
Figure BDA0003769902760000081
The class activation graph function is displayed in an overlaying mode with the CWD time-frequency analysis feature graph in the form of thermodynamic diagram, and the CWD feature graph of the high sensitivity area is analyzed and intercepted;
inputting the CWD characteristic diagram of the high-sensitivity region into a support vector machine for identification; the method is mainly applied to the individual identification scene of the radiation source under low computation and high signal-to-noise ratio.
Other steps and parameters are the same as those in one of the first to fifth embodiments.
The sixth specific implementation mode: this embodiment differs from the first embodiment in that the tagged non-pre-distorted different radiation source signal v in the first stepin(n) input to a predistorter which outputs tagged predistorted different radiation source signals vpd(n); the specific process is as follows:
and (4) carrying out linearization treatment on the power amplifier based on a QRD-LS algorithm to obtain a radiation source predistortion signal.
Other steps and parameters are the same as those in the first embodiment.
The seventh concrete implementation mode: in this embodiment, the difference between the first embodiment and the sixth embodiment is that the pre-distorter outputs the signed pre-distortional source-difference signal v in the first steppd(n) input to a Power Amplifier (PA) which outputs a tagged signal vpa(n); the specific expression is as follows:
Figure BDA0003769902760000082
wherein v ispd(n) is the predistorter output signal, vpa(n) is the output signal of the power amplifier, n is the signal index, and K is the MP model order; q is the MP model depth; h iskqIs a memory polynomial coefficient;
the MP model is a Memory Polynomial model (MP model);
the signal v with label output by the power amplifier in the first steppa(n) input
Figure BDA0003769902760000083
The output signal is input into the digital predistortion trainer, and the output of the digital predistortion trainer
Figure BDA0003769902760000084
G is the power amplifier gain; the specific expression is as follows:
Figure BDA0003769902760000085
wherein the content of the first and second substances,
Figure BDA0003769902760000086
for the output signal of the digital predistortion trainer, vpaOutputting a signal for the power amplifier, wherein n is a signal index, and K is the order of the MP model; q is the depth of the MP model; g is the power amplifier gain; w is akqAre digital predistorter coefficients.
Digital predistorter coefficients wkqIs composed of
Figure BDA0003769902760000091
The solution of (1).
Other steps and parameters are the same as those in the sixth embodiment.
The specific implementation mode is eight: the seventh embodiment is different from the seventh embodiment in that, in the fifth step, the sincent network sequentially includes: the device comprises an input layer, a filter structure, a first pooling layer, a first convolution layer, a second pooling layer, a second convolution layer, a third pooling layer, a first full-link layer, a second full-link layer, a third full-link layer, a fourth full-link layer, a LeakyReLU activation function and a softmax layer.
Other steps and parameters are the same as those in the seventh embodiment.
The specific implementation method nine: the difference between this embodiment and the eighth embodiment is that the filter structure is an 80-dimensional filter structure, the initial bandwidth of each filter is about 0.37MHz, and the initial frequency and the cutoff frequency of the filter structure are continuously distributed from 0 to 30 MHz.
The other steps and parameters are the same as those in the eighth embodiment.
The specific implementation mode is ten: the present embodiment is different from the ninth embodiment in that the first pooling layer is a pooling layer having a window size of 3;
the convolution kernel sizes of the first convolution layer and the second convolution layer are both 5 multiplied by 5;
the second and third pooling layers are pooling layers with a window size of 3;
and a BN layer is connected behind each full connecting layer in the first full connecting layer, the second full connecting layer, the third full connecting layer and the fourth full connecting layer.
Other steps and parameters are the same as those in the ninth embodiment.
As shown in fig. 2, the sincenet filter structure is substantially a filter extraction of more effective low-dimensional features at the first layer of the network, and realizes the extraction of frequency sensitivity regions, which perform sensitivity analysis with the class activation diagrams and select the behaviors of the sensitivity regions, with the consistent purpose. At a sampling rate of 60MHz, a first layer SincNet of the network is a filter structure with 80 dimensions, the initial bandwidth of each filter is about 0.37MHz, and the initial frequency and the cut-off frequency of the structure are continuously distributed from 0 to 30 MHz. On the basis, dimensionality reduction is carried out by using a pooling layer with the window size of 3, the convolution layers with the convolution kernel size of 5 multiplied by 5 and the pooling layer with the window size of 3 are output, the bottom layer characteristics of the sample are embedded, and the four full-connection layers are output. And a BN layer is arranged behind the full connection layer, a LeakyReLU activation function is adopted, a Dropout ratio is set to be 0.5, and finally the individual signal identification of the radiation source after pre-distortion is completed after normalization by a softmax layer. The method is mainly applied to scenes of individual identification of radiation sources with more severe signal-to-noise ratio.
The following examples were used to demonstrate the beneficial effects of the present invention:
the first embodiment is as follows:
the characteristic image obtained by performing CWD conversion on the two pre-distorted radiation source individuals is input into the obtained ResNet50-CWD network, the weight of the characteristic image classification is reversely propagated back to the input time spectrogram through an activation function, and finally the class activation image of the trained network is obtained through Grad-CAM, as shown in FIGS. 3a, 3b, 3c and 3d, the bright area in FIGS. 3a, 3b, 3c and 3d represents the part with large classification weight of the region, and plays an important role in network identification. It can be found that the bright areas are mainly distributed along the time-frequency part of the main signal of the time-frequency diagram, while the dark areas are mainly distributed on the noise part of the non-main frequency, and the Grad-CAM method has practical physical significance for analyzing the sensitive area of the time-frequency diagram. The distribution of the sensitive areas of 500 samples of each of the three radiation source individuals is observed and analyzed, and the sensitive areas of different samples of the same individual are converged, while the distribution positions of the sensitive areas of the three radiation source individuals are obviously different, for example, SE1 and SE2, the main sensitive area of the class activation diagram of SE1 is distributed at the edge part of modulation, and SE2 is mainly distributed at the intermediate frequency part.
The characteristics of the local information of the sensitive region are analyzed secondarily, 1500 corresponding regions of characteristic patterns are intercepted aiming at the sensitive region of the class activation graph of the SE1, and the class separability of the local characteristics of the radiation source signals at the sensitive region after the predistortion is discussed, as shown in FIGS. 4a, 4b, 4c, 4d, 4e and 4 f. And summing and averaging channels of 1500 sample sensitive areas of the three radiation sources, and drawing a time-frequency image of two GB channels. It can be found that the sensitive areas differ significantly for different radiation sources, especially in the B channel. Overall, the inter-class dispersion is stronger.
And the high-precision identification is directly carried out by utilizing the sensitive area of the characteristic diagram. An SVM is adopted as a classifier, SE1, SE2 and SE3 are recognition objects, and the number of samples after pre-distortion is 4500. Comparing the recognition accuracy of the traditional characteristics under the SVM, as shown in the table 1. The result shows that the high-sensitivity region identification can obtain the identification precision equivalent to that of the full region identification at a high signal-to-noise ratio, but the identification precision is reduced sharply at a low signal-to-noise ratio, and the identification precision is reduced by 9.2 percent at-5 db. This result is reasonable. Since the high sensitivity region of this section is a rectangular window that is manually cut out by observation after analysis, and is not completely correct, the actual real sensitive region should be an irregular, discrete region in practice. On the other hand, with high-sensitivity region identification, the calculation amount can be necessarily reduced, which is a great advantage of the method.
TABLE 1 recognition accuracy with SVM under different characteristics
Figure BDA0003769902760000101
Example two:
the method comprises the steps of identifying a signal time sequence after input digital predistortion by using a neural network based on a SincNet structure, training single pulse 1000 sampling points under 1500 samples by using 500 samples of SE1, SE2 and SE3, and testing by using pulse 1000 sampling points of the same 1500 samples in different time periods. And setting a self-adaptive learning rate function, setting input 128 samples by the batch, and setting 1000 sampling points of each sample to finally obtain a filter frequency band learned by the first layer of SincNet, namely the signal sensitive frequency. The network performance constructed herein is verified, and the recognition accuracy under different signal-to-noise ratios and the results of comparing other networks are finally obtained as shown in table 2. Compared with the most accurate ResNeXt50-CWD in the comparison method, the SincNet network added with the predistortion has the advantages of greatly improving the identification accuracy in individual identification of the radiation source under strong noise, improving the accuracy by 2.5% under-5 db, and ensuring that the identification accuracy is equivalent under high signal-to-noise ratio. Because the structure of the SincNet filter realizes accurate selection of the sensitive area, the interference of noise can be better filtered, and the result is better under the condition of low signal-to-noise ratio. The total number of floating-point arithmetic operations of the algorithm and the total number of parameters of the algorithm model are compared, and the result is shown in table 3. From the table, the ResNeXt50-CWD method has large computation amount, which is caused by the complex network structure, and the computation amount and parameters of the method based on the SincNet network are reduced by half, which is because the network directly filters in the first layer, learns the filter parameters, pays more attention to the sensitive area and reduces the network complexity.
TABLE 2 recognition accuracy of different networks at different SNR
Figure BDA0003769902760000111
TABLE 3 comparison of the calculation amounts
Figure BDA0003769902760000112
The present invention is capable of other embodiments and its several details are capable of modifications in various obvious respects, all without departing from the spirit and scope of the present invention.

Claims (10)

1. The individual identification method of the radiation source based on the class activation graph and the SincNet network is characterized in that: the method comprises the following specific processes:
judging whether the radiation source signal to be detected is a high signal-to-noise ratio radiation source signal or a low signal-to-noise ratio radiation source signal, and executing the step A to the step D if the radiation source signal to be detected is the high signal-to-noise ratio radiation source signal; if the radiation source signal to be detected is a low signal-to-noise ratio radiation source signal, executing the first step to the third step;
the high signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio of more than 15 db;
the low signal-to-noise ratio radiation source signal is a radiation source signal with a signal-to-noise ratio of less than or equal to 15 db;
the specific process is as follows:
if the radiation source signal to be detected is a high signal-to-noise ratio signal, the specific steps are as follows:
step A, carrying out label-carrying non-predistortion different radiation source signals vin(n) inputting predistorters which output tagged predistorted differential radiation source signals vpd(n)；
Output signal v of predistorter with labeled predistortion different radiation sourcepd(n) input to a power amplifier, which outputs a tagged signal vpa(n)；
And B: for the tagged signal v output by the power amplifierpa(n) performing CWD processing, and extracting a marked CWD time-frequency distribution characteristic diagram;
and C: inputting the marked CWD time-frequency distribution characteristic diagram extracted in the step B into a ResNet50 network for training to obtain a trained ResNet50 network;
step D: inputting an un-predistorted radiation source signal to be tested into a predistorter, and outputting a predistorted radiation source signal by the predistorter; a predistortion radiation source signal output by the predistorter is input into a power amplifier, and the power amplifier outputs a signal; performing CWD processing on the signal output by the power amplifier, and extracting a CWD time-frequency distribution characteristic diagram;
inputting the extracted CWD time-frequency distribution characteristic diagram into a trained ResNet50 network, and analyzing a class activation diagram to obtain a class activation diagram function;
the class activation graph function is displayed in an overlaying mode with the CWD time-frequency analysis characteristic graph in a thermodynamic diagram mode, and the CWD characteristic graph of the sensitive area is analyzed and intercepted;
inputting a CWD characteristic diagram of the sensitive area into a support vector machine for identification;
if the radiation source signal to be detected is a low signal-to-noise ratio signal, the specific steps are as follows:
the method comprises the following steps: labeled non-pre-distorted different radiation source signals vin(n) input to a predistorter which outputs tagged predistorted different radiation source signals vpd(n)；
Output signal v of predistorter with labeled predistortion different radiation sourcepd(n) input to a power amplifier, which outputs a tagged signal vpa(n)；
Tagged signal v output by predistorterpd(n) an input delay;
tagged signal v output by a power amplifierpa(n) input
Figure FDA0003769902750000021
The output signal is input into the digital predistortion trainer, and the output of the digital predistortion trainer
Figure FDA0003769902750000022
G is the power amplifier gain;
output signal of digital predistortion trainer
Figure FDA0003769902750000023
And the delay output signal vpd(n) carrying out subtraction operation to output an error e (n), training the model parameters of the digital predistortion trainer by the adaptive algorithm, and when the model of the digital predistortion trainer happens to be correctWhen it is an inverse model of a power amplifier, there are
Figure FDA0003769902750000024
Obtaining a corresponding trained digital predistortion trainer after the training is finished; otherwise, the error e (n) continues to train the model parameters of the digital predistortion trainer through the adaptive algorithm until the model parameters of the digital predistortion trainer are trained
Figure FDA0003769902750000025
Step two: inputting the different radiation source signals with labels without predistortion into a predistorter, and outputting the different radiation source signals with labels with predistortion by the predistorter;
the pre-distorter outputs different pre-distorter radiation source signals with tags to the power amplifier, and the power amplifier outputs the signals with the tags;
tagged signal v output by a power amplifierpa(n) input
Figure FDA0003769902750000026
The output signal is input into a trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; g is the power amplifier gain;
inputting the output signal of the trained digital predistortion trainer into a SincNet network for training to obtain the trained SincNet network;
step three: inputting an un-predistorted radiation source signal to be tested into a predistorter, and outputting a predistorted radiation source signal by the predistorter; inputting a predistortion radiation source signal output by a predistorter into a power amplifier, and outputting a signal to be tested by the power amplifier;
signal input to be tested output by power amplifier
Figure FDA0003769902750000027
The output signal is input into a trained digital predistortion trainer, and the trained digital predistortion trainer outputs a signal; g is the power amplifier gain;
and inputting the output signal of the trained digital predistortion trainer into the trained SincNet network to finish the identification of the individual signal of the radiation source after predistortion.
2. The individual identification method of a radiation source based on a class activation graph and a SincNet network according to claim 1, wherein: the labeled non-pre-distorted different radiation source signals v in the step Ain(n) input to a predistorter which outputs tagged predistorted different radiation source signals vpd(n); the specific process is as follows:
and (4) carrying out linearization treatment on the power amplifier based on a QRD-LS algorithm to obtain a radiation source predistortion signal.
3. The individual identification method of a radiation source based on a class activation graph and a SincNet network according to claim 2, wherein: the pre-distorted different radiation source signals v with labels output by the pre-distorter in the step Apd(n) input to a power amplifier, which outputs a tagged signal vpa(n); the specific expression is as follows:
Figure FDA0003769902750000031
wherein v ispd(n) is the predistorter output signal, vpa(n) is the output signal of the power amplifier, n is the signal index, and K is the MP model order; q is the MP model depth; h iskqIs a memory polynomial coefficient;
the MP model is a memory polynomial model.
4. The individual identification method of a radiation source based on a class activation graph and a SincNet network according to claim 3, wherein: the signal v with label output by the power amplifier in the step Bpa(n) performing CWD processing, and extracting a marked CWD time-frequency distribution characteristic diagram; the specific process is as follows:
the CWD distribution is defined as
Figure FDA0003769902750000032
Wherein CWD (t, f) is an output time-frequency distribution characteristic diagram, t is time, f is frequency, sigma is a scale factor, tau is a time shift parameter, and v is a power amplifier output signal vpa(n)，v*For outputting a signal v to a power amplifierpa(n) and j represents an imaginary number.
5. The individual identification method of the radiation source based on the class activation graph and the SincNet network according to claim 4, wherein: in the step D, inputting the radiation source signal to be tested without predistortion into a predistorter, and outputting the radiation source signal by the predistorter; a predistortion radiation source signal output by the predistorter is input into a power amplifier, and the power amplifier outputs a signal; performing CWD processing on the signal output by the power amplifier, and extracting a CWD time-frequency distribution characteristic diagram;
inputting the extracted CWD time-frequency distribution characteristic diagram into a trained ResNet50 network, and analyzing a class activation diagram to obtain a class activation diagram function;
the class activation graph function is displayed in an overlaying mode with the CWD time-frequency analysis feature graph in the form of thermodynamic diagram, and the CWD feature graph of the sensitive area is analyzed and intercepted;
inputting the CWD characteristic diagram of the high sensitivity area into a support vector machine for identification;
the specific process is as follows:
inputting an un-predistorted radiation source signal to be tested into a predistorter, and outputting a predistorted radiation source signal by the predistorter; inputting a predistortion radiation source signal output by a predistorter into a power amplifier, and outputting a signal by the power amplifier; performing CWD processing on the signal output by the power amplifier, and extracting a CWD time-frequency distribution characteristic diagram;
inputting the extracted CWD time-frequency distribution characteristic diagram into a trained ResNet50 network;
let k feature maps generated by the last convolutional layer in the ResNet50 network be
Figure FDA0003769902750000041
Wherein W and b are the weight and learning bias of the classification layer neurons, respectively; f. ofactA non-linear activation function for the network;
Figure FDA0003769902750000042
is the product of the tensors, xcThe extracted CWD time-frequency distribution characteristic graph is obtained;
characteristic diagram AkOutput through ResNet50 network output layer as
Figure FDA0003769902750000043
Wherein the content of the first and second substances,
Figure FDA0003769902750000044
c-type learning weight corresponding to the ResNet50 network output layer unit; y iscA score value with an output category of c is obtained;
defining the activation value of the kth cell of the last convolutional layer in the ResNet50 network at position (i, j)
Figure FDA0003769902750000045
The weight corresponding to the object class c is
Figure FDA0003769902750000046
Wherein Z represents the total number of pixels;
setting a ReLU activation function after the ResNet50 network output layer, and obtaining a sample x through the ReLU activation functioncClass activation graph function through ResNet50 network
Figure FDA0003769902750000047
The class activation graph function is displayed in an overlaying mode with the CWD time-frequency analysis characteristic graph in a thermodynamic diagram mode, and the CWD characteristic graph of the sensitive area is analyzed and intercepted;
inputting the CWD characteristic diagram of the sensitivity region into a support vector machine for identification.
6. The individual identification method of a radiation source based on a class activation graph and a SincNet network according to claim 5, wherein: the labeled non-pre-distorted different radiation source signals v in the step onein(n) input to a predistorter which outputs tagged predistorted different radiation source signals vpd(n); the specific process is as follows:
and (4) carrying out linearization treatment on the power amplifier based on a QRD-LS algorithm to obtain a radiation source predistortion signal.
7. The individual identification method of a radiation source based on a class activation graph and a SincNet network according to claim 6, wherein: the pre-distorts the different radiation source signals v with labels output by the pre-distorter in the first steppd(n) input to a power amplifier, which outputs a tagged signal vpa(n); the specific expression is as follows:
Figure FDA0003769902750000051
wherein v ispd(n) is the predistorter output signal, vpa(n) is the output signal of the power amplifier, n is the signal index, and K is the MP model order; q is the MP model depth; h iskqIs a memory polynomial coefficient;
the MP model is a memory polynomial model;
the signal v with label output by the power amplifier in the first steppa(n) input
Figure FDA0003769902750000052
The output signal is input into the digital predistortion trainer, and the output of the digital predistortion trainer
Figure FDA0003769902750000053
G is the power amplifier gain; the specific expression is as follows:
Figure FDA0003769902750000054
wherein, the first and the second end of the pipe are connected with each other,
Figure FDA0003769902750000055
for the output signal of the digital predistortion trainer, vpaOutputting a signal for the power amplifier, wherein n is a signal index, and K is the order of the MP model; q is the depth of the MP model; g is the power amplifier gain; w is akqAre digital predistorter coefficients.
8. The individual identification method of a radiation source based on a class activation graph and a SincNet network according to claim 7, wherein: the SincNet network in the fifth step comprises in sequence: the device comprises an input layer, a filter structure, a first pooling layer, a first convolution layer, a second pooling layer, a second convolution layer, a third pooling layer, a first full-link layer, a second full-link layer, a third full-link layer, a fourth full-link layer, a LeakyReLU activation function and a softmax layer.
9. The individual identification method of the radiation source based on the class activation graph and the SincNet network according to claim 8, wherein: the filter structure is an 80-dimensional filter structure, the initial bandwidth of each filter is about 0.37MHz, and the initial frequency and the cut-off frequency of the filter structure are continuously distributed from 0 to 30 MHz.
10. The individual identification method of the radiation source based on the class activation graph and the SincNet network according to claim 9, wherein: the first pooling layer is a pooling layer with a window size of 3;
the convolution kernels of the first convolution layer and the second convolution layer are 5 multiplied by 5;
the second and third pooling layers are pooling layers with a window size of 3;
and a BN layer is connected behind each full connecting layer in the first full connecting layer, the second full connecting layer, the third full connecting layer and the fourth full connecting layer.
CN202210903562.1A 2022-07-28 2022-07-28 Radiation source individual identification method based on class activation diagram and SincNet network Pending CN115270878A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210903562.1A CN115270878A (en) 2022-07-28 2022-07-28 Radiation source individual identification method based on class activation diagram and SincNet network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210903562.1A CN115270878A (en) 2022-07-28 2022-07-28 Radiation source individual identification method based on class activation diagram and SincNet network

Publications (1)

Publication Number Publication Date
CN115270878A true CN115270878A (en) 2022-11-01

Family

ID=83770500

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210903562.1A Pending CN115270878A (en) 2022-07-28 2022-07-28 Radiation source individual identification method based on class activation diagram and SincNet network

Country Status (1)

Country Link
CN (1) CN115270878A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116401588A (en) * 2023-06-08 2023-07-07 西南交通大学 Radiation source individual analysis method and device based on deep network

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116401588A (en) * 2023-06-08 2023-07-07 西南交通大学 Radiation source individual analysis method and device based on deep network
CN116401588B (en) * 2023-06-08 2023-08-15 西南交通大学 Radiation source individual analysis method and device based on deep network

Similar Documents

Publication Publication Date Title
CN108573225B (en) Partial discharge signal pattern recognition method and system
CN110197205B (en) Image identification method of multi-feature-source residual error network
CN105844279A (en) Depth learning and SIFT feature-based SAR image change detection method
CN110610717B (en) Separation method of mixed signals in complex frequency spectrum environment
CN107729926B (en) Data amplification method and machine identification system based on high-dimensional space transformation
CN109085469A (en) A kind of method and system of the signal type of the signal of cable local discharge for identification
CN107491793B (en) Polarized SAR image classification method based on sparse scattering complete convolution
CN109993236A (en) Few sample language of the Manchus matching process based on one-shot Siamese convolutional neural networks
CN110824586B (en) Rainfall prediction method based on improved decision tree algorithm
CN109711314B (en) Radar radiation source signal classification method based on feature fusion and SAE
CN111583230B (en) Low-rank sparse decomposition hyperspectral abnormal target detection method based on local features
CN112949387A (en) Intelligent anti-interference target detection method based on transfer learning
CN115270878A (en) Radiation source individual identification method based on class activation diagram and SincNet network
CN115270872A (en) Radar radiation source individual small sample learning and identifying method, system, device and medium
CN115345909B (en) Hyperspectral target tracking method based on depth space spectrum convolution fusion characteristics
CN113780242A (en) Cross-scene underwater sound target classification method based on model transfer learning
CN113962264A (en) Aero-engine rotor system fault diagnosis algorithm based on deep learning
CN114520758A (en) Signal modulation identification method based on instantaneous characteristics
CN116312860A (en) Agricultural product soluble solid matter prediction method based on supervised transfer learning
CN115061094A (en) Radar target identification method based on neural network and SVM
CN110619366A (en) Neural network-based fungus MALDI-TOF mass spectrum data identification method
CN115169406A (en) Instantaneous phase fingerprint feature enhancement method based on empirical mode decomposition
CN115205701A (en) Method for extracting sample based on rice phenological characteristics
CN113343796A (en) Knowledge distillation-based radar signal modulation mode identification method
CN115186717A (en) Radiation source individual identification method under power amplifier digital predistortion

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination