CN110532932B - Method for identifying multi-component radar signal intra-pulse modulation mode - Google Patents
Method for identifying multi-component radar signal intra-pulse modulation mode Download PDFInfo
- Publication number
- CN110532932B CN110532932B CN201910787759.1A CN201910787759A CN110532932B CN 110532932 B CN110532932 B CN 110532932B CN 201910787759 A CN201910787759 A CN 201910787759A CN 110532932 B CN110532932 B CN 110532932B
- Authority
- CN
- China
- Prior art keywords
- network
- time
- signal
- component
- signals
- 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.)
- Active
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
- G01S13/88—Radar or analogous systems specially adapted for specific applications
- G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Remote Sensing (AREA)
- Radar, Positioning & Navigation (AREA)
- Physics & Mathematics (AREA)
- Computer Networks & Wireless Communication (AREA)
- General Physics & Mathematics (AREA)
- Electromagnetism (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention relates to the field of automatic recognition algorithms for deep learning, in particular to a method for recognizing an intra-pulse modulation mode of a multi-component radar signal. Acquiring time-frequency images of single-component or overlapped multi-component radar signals of several different intra-pulse modulation modes; preprocessing a radar signal time-frequency image by using an image processing algorithm, and making a training set and a test set by using signal types contained in radar signals as labels; designing a pre-training network based on a convolutional neural network to extract radar signal time-frequency image features, and designing a multi-component signal classification network based on reinforcement learning to obtain a classification recognition result; training, testing and perfecting network structure and parameters; and realizing classification identification of the multi-component signals. The multi-component radar signal identification algorithm has wide radar signal type adaptation range and high identification accuracy under the condition of low signal-to-noise ratio, and realizes the intra-pulse modulation mode identification of random overlapping multi-component radar signals.
Description
Technical Field
The invention relates to the field of automatic identification algorithms for deep learning, in particular to a method for identifying an intra-pulse modulation mode of a multi-component radar signal.
Background
Identification of the radar signal intra-pulse modulation mode is an important link in modern electronic information reconnaissance and electronic support systems.
Because radar signal density continuously increases in modern electronic warfare environment, modern radar signal adopts the pulse compression signal of big hour width moreover mostly, and the radar reconnaissance system often can intercept the pulse of time domain overlapping, forms the multicomponent radar signal. Most of the existing radar signal modulation mode identification technologies have no adaptability to multi-component signal environments, and signal identification errors or identification failures are caused. Therefore, analysis and processing for multi-component signals are issues that are urgently needed to be solved in current radar reconnaissance systems.
At present, the identification method of the multi-component radar signal has two ideas: one idea is to use the separability of the signal in some transform domains to extract the features of the signal in the transform domains to classify and identify the multi-component signal; the other idea is to use a separation method of multi-component signals to combine with a recognition method of single-component signals for classification and recognition. The remainder and the even in 2012 put forward a multi-component signal identification method for a phase shift keying signal by utilizing the separability of the phase shift keying signal on a cyclic frequency axis, and when the signal-to-noise ratio is 0dB, the average correct identification rate for 3-class phase shift keying radar signals reaches 97%. A multi-component radar signal identification method combining independent component analysis and wavelet transformation is proposed in 2016, and the average correct identification rate of 4 types of radar signals reaches over 90% when the signal-to-noise ratio is 0 dB.
The identification methods of the multi-component radar signals proposed at present have some problems: the identification method based on the signal transform domain feature extraction has limitations, is effective only for a certain specific signal, and is difficult to adapt to a wide range of radar signal types; the identification effect of the identification method based on the multi-component signal separation is determined by the separation effect of the multi-component signal separation algorithm to a great extent, however, the currently proposed separation algorithm has the problems of poor anti-noise performance, high calculation amount and the like, and the problem of multi-component signal separation of time-frequency domain overlapping is difficult to solve, and the problems limit the identification capability of the algorithm.
Disclosure of Invention
The invention aims to provide a multi-component radar signal intra-pulse modulation mode identification method, which realizes intra-pulse modulation mode identification of typical 8-class randomly overlapped radar signals under the condition of low signal-to-noise ratio, and has adaptability to intra-pulse modulation mode identification of single-component radar signals.
The embodiment of the invention provides a method for identifying an intra-pulse modulation mode of a multi-component radar signal, which comprises the following steps:
the method comprises the following steps: acquiring time-frequency images of single-component or overlapped multi-component radar signals of different intra-pulse modulation modes, wherein the time-frequency images comprise LFM signals, MP signals, SFM signals, BPSK signals, 2FSK signals, 4FSK signals, EQFM signals and Frank signals, the signals are used as sample signals, and signals received by a radar receiver are converted by using time-frequency distribution to obtain radar signal time-frequency images;
step two: preprocessing the radar signal time-frequency image obtained in the first step by using an image processing algorithm, inhibiting noise of the time-frequency image through two-dimensional wiener filtering, adjusting the size of the time-frequency image and normalizing the amplitude, taking the preprocessed radar signal time-frequency image as a sample, taking the signal type contained in the radar signal as a label, and making a training set and a test set;
step three: according to algorithm requirements, the whole network extracts time-frequency image features and carries out multi-label classification identification according to the extracted features, a pre-training network based on a convolutional neural network is designed to extract radar signal time-frequency image features, and a multi-component signal classification network based on reinforcement learning is designed to obtain classification identification results;
step four: training the network architecture designed in the third step according to the training set obtained in the second step, testing the network identification effect by using the test set obtained in the second step, and perfecting the network structure and parameters according to the test result to obtain a trained network;
step five: any random overlapping signal in the first step is processed in the second step and then serves as a radar signal time-frequency image sample to be identified, the radar signal time-frequency image sample is input into the network trained in the fourth step, and the trained network can provide a radar signal type set contained in the current signal according to input, so that classification and identification of multi-component signals are achieved;
the invention also includes such structural features:
the first step comprises the following steps:
the specific method of the time-frequency distribution comprises the following steps:
for a received signal x (t), cohen type time frequency distribution is adopted, and the mathematical expression is as follows:
in the above formula, t and ω represent independent variable time and angular frequency of time-frequency distribution, and Φ (τ, v) is referred to as kernel function, according to distribution characteristics of radar signals and cross terms in a fuzzy domain, aiming at signals with different frequency modulation slopes, the invention provides that two kernel functions are used for respectively obtaining time-frequency distribution of radar signals, and the expression of the kernel function is respectively:
in the above formula, α, β, γ and e are used to adjust the size of the kernel function, and the Cohen-type time-frequency distribution images of the radar signal under the two kernel functions are respectively obtained and are jointly used as the time-frequency distribution information of the current radar signal;
in the third step, the specific method for designing the network architecture is as follows:
a) Designing a convolutional neural network architecture for feature extraction, wherein the convolutional neural network architecture consists of 3 convolutional layers and 3 pooling layers, and vectorizing the output of the last layer of the network to obtain a feature vector of the current time-frequency image;
b) Taking a structure of a full connection layer and a Softmax layer as a classification network unit, constructing a plurality of classification network units to respectively correspond to label outputs of multi-component signals, classifying the characteristic vectors obtained in the step a), and taking the characteristic vectors and a convolutional neural network framework together as a pre-training time-frequency image characteristic extraction network;
c) Designing a deep reinforcement learning cyclic neural network for multi-component signal classification and identification, replacing a classification network unit in the step b), taking the feature vector obtained in the step a) and the classification historical record of the current cyclic step as input together, and outputting a multi-component signal identification result through a multi-cycle iterative classification process;
in the fourth step, the specific method for training and testing is as follows:
d) After the second step, the radar signal time-frequency image subjected to image preprocessing is used as the input of a convolutional neural network, each classification network unit outputs a classification label corresponding to one radar signal component, the time-frequency image characteristic extraction network is pre-trained, and the parameters of the convolutional neural network are reserved after training;
e) Taking the convolutional neural network and the deep reinforcement learning cyclic neural network obtained by training in the step d) as new classification networks, and in order to improve the network training efficiency and keep the model parameters of the convolutional neural network unchanged, independently training the model parameters of the deep reinforcement learning cyclic neural network by using the same training set as that in the step d), and keeping all the parameters of the final convolutional neural network and the deep reinforcement learning cyclic neural network;
f) Samples of the test set are input into a classifier consisting of a convolutional neural network and a deep reinforcement learning cyclic neural network one by one, the classifier can output a signal type set contained in the current sample for many times in a circulating manner, the samples are compared with a standard signal type set, and finally the integral correct recognition probability and the recognition accuracy and the recall rate of each signal type are obtained through statistical calculation;
g) Perfecting a network structure and parameters according to the recognition effect test result of the step f), training a new network structure, repeating the network adjustment and training process until the recognition effect test result of the network reaches an expected value, and finishing the training;
the invention has the beneficial effects that:
1. the time-frequency analysis method provided by the invention designs different kernel functions according to the characteristics of different radar signals, gives consideration to the energy distribution conditions of the radar signals with different frequency modulation slopes, and obtains a time-frequency image with higher signal energy time-frequency aggregation;
2. according to the method, the convolutional neural network is pre-trained, deep features of a radar signal time-frequency image are obtained, and meanwhile the training efficiency of a subsequent multi-component signal classification network is improved;
3. the multi-component classification network is trained by using a reinforcement learning training method, so that the adaptability of the network to time-frequency image samples is improved; and the relation among the multiple classification recognition results is established by utilizing a circulating network architecture, so that the accuracy and the reliability of the classification recognition results are improved.
Drawings
FIG. 1 is a flow chart of a method for identifying an intra-pulse modulation mode of a multi-component radar signal;
FIG. 2 is a structural diagram of a pre-training time-frequency image feature extraction network based on a convolutional neural network;
FIG. 3 is a block diagram of a multi-component signal classification identification network based on convolutional neural network and reinforcement learning according to the present invention;
FIG. 4 is a schematic diagram of the relationship between the average correct recognition rate and the signal-to-noise ratio of a multi-component radar signal according to the present invention;
FIG. 5 is a schematic diagram of the relationship between the recognition accuracy and the signal-to-noise ratio of 8-class randomly overlapped multi-component radar signals according to the present invention;
FIG. 6 is a schematic diagram of the relationship between the recognition recall ratio and the signal-to-noise ratio of 8-type randomly overlapped multi-component radar signals according to the present invention;
FIG. 7 is a diagram illustrating the relationship between the average correct recognition rate and the signal-to-noise ratio of a single-component radar signal according to the present invention.
Detailed Description
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, the present invention is further described with reference to the accompanying drawings:
the technical scheme of the invention is realized as follows:
fig. 1 is a flowchart of an identification algorithm of an intra-pulse modulation scheme of a multi-component radar signal according to the present invention, and the steps and the principle of the algorithm will be described in detail with reference to fig. 1.
Step 1: the method comprises the steps of obtaining time-frequency images of single-component or overlapped multi-component radar signals of different intra-pulse modulation modes, wherein the time-frequency images comprise LFM signals, MP signals, SFM signals, BPSK signals, 2FSK signals, 4FSK signals, EQFM signals and Frank signals, taking the signals as sample signals, and converting signals received by a radar receiver into time-frequency images by using time-frequency distribution.
In this step, the mathematical model of the radar signal after the noise is added can be written as:
x(t)＝s(t)+n(t) (1)
in the above formula, x (t) is a signal received by the radar, s (t) is a signal, and n (t) represents channel noise.
The invention adopts Cohen-like time-frequency distribution for converting radar signals into time-frequency images
Where t and ω represent the time and angular frequency of the time-frequency distribution, and φ (τ, v) is called the kernel function. According to the distribution characteristics of radar signals and cross terms in a fuzzy domain, aiming at signals with different frequency modulation slopes, the invention provides that two kernel functions are utilized to respectively obtain the time-frequency distribution of the radar signals, and the expression of the kernel functions is respectively
Wherein the kernel parameters α, β, γ and ε are used to adjust the size of the kernel. Aiming at radar signals with lower frequency modulation slope, the invention adopts a double Gaussian kernel function phi 1 (τ, v), regulatory nucleusThe long axis of the function parameter points to the time delay tau axis to obtain the main energy of the low frequency modulation slope signal; aiming at radar signals with higher frequency modulation slope, the invention provides a new kernel function phi 2 (tau, v), adjusting parameters of the kernel function to enable the long axis of the kernel function to point to the frequency shift v axis, and acquiring main energy of the high frequency shift slope signal. Cohen time frequency distribution images of the radar signals under the two kernel functions are respectively obtained and are jointly used as time frequency distribution information of the current radar signals.
Step 2: preprocessing a radar signal time-frequency image by using an image processing algorithm, inhibiting time-frequency image noise through two-dimensional wiener filtering, adjusting the size of the image and normalizing the amplitude, and then manufacturing a training set and a test set according to the processed radar signal time-frequency image and a signal class label set contained in a corresponding signal.
The algorithm adopts two-dimensional wiener filtering to inhibit the noise of the time-frequency image, and adjusts the amplitude of a certain pixel point of the time-frequency image by calculating the mean value and the variance in the filtering neighborhood of the pixel point, so as to realize the two-dimensional wiener filtering of the time-frequency image; adjusting the size of the time-frequency image by adopting a bilinear interpolation method, and calculating the amplitude of each pixel point of the image after the size is adjusted by adopting a line-by-line linear interpolation method to realize the size adjustment of the time-frequency image; in order to facilitate the work of a subsequent classifier, the amplitude of the time-frequency image is divided by the maximum amplitude value to realize amplitude normalization; and adding a label to the preprocessed time-frequency image to obtain a data set. In particular, the output forms of the Softmax layer and the deep reinforcement learning recurrent neural network are different, and the label forms added by the time-frequency images are also different: the label form corresponding to the output of the Softmax layer is splicing of the label vectors of the one-hot codes of the signal components, and the label form corresponding to the output of the deep reinforcement learning recurrent neural network is the number of the signal components of the signal types contained in the current signal.
And 3, step 3: and designing a multi-component radar signal time-frequency image identification network.
The internal network architecture of each module of the pre-training feature extraction network and the multi-component signal classification and identification network designed by the algorithm is shown in fig. 2 and 3.
And (3.1) designing a convolutional neural network for extracting the time-frequency image characteristics, inputting the preprocessed time-frequency image into the convolutional neural network, sequentially passing through 3 convolutional layers and then connecting with a network structure of a pooling layer, and vectorizing the output of the last pooling layer to obtain the characteristic vector of the time-frequency image. The specific operation of the convolutional layer is: traversing the input image by taking a convolution kernel as a basic unit to carry out multiply-add operation to obtain an output characteristic image; the concrete operation of the pooling layer is as follows: in a region, a specific value is taken as an output value, and the pooling mode in the algorithm adopts average pooling, namely, the average value of the region is taken as the output value; the vectorization process is to spread the multi-dimensional matrix features into feature vectors.
And (3.2) taking a structure of a full connection layer and a Softmax layer as a classification network unit, establishing a mapping relation between the feature vectors and the signal types through the full connection layer, mapping the output of the full connection layer into probability values of the signal types with the value range of [0,1] by using the Softmax layer, outputting the signal types corresponding to the highest probabilities as labels of the classification network unit, constructing label outputs of a plurality of classification network units corresponding to the multi-component signals respectively, classifying the feature vectors obtained in the step (3.1), and taking the feature vectors and the convolutional neural network framework as a pre-training time-frequency image feature extraction network together.
And (3.3) designing a deep reinforcement learning cyclic neural network for multi-component signal classification and identification, replacing the classification network unit in the step (3.2), and outputting a multi-component signal identification result through a multi-cycle iterative classification process. In order to improve the accuracy of the network classification result, the input of the network comprises the number of signal components of each signal type output by the recurrent neural network before the current recurrent step, namely the classification history of the recurrent neural network, besides the feature vector output by the recurrent neural network.
And 4, step 4: and training a multi-component radar signal time-frequency image recognition network.
(4.1) after the step 2, taking the radar signal time-frequency image subjected to image preprocessing as the input of a convolutional neural network, outputting a classification label corresponding to a radar signal component by each classification network unit, calculating a square error loss function of the output label and a standard label, pre-training the time-frequency image feature extraction network by using an error back propagation and random gradient descent algorithm, and keeping parameters of the convolutional neural network after training.
And (4.2) taking the convolutional neural network and the deep reinforcement learning cyclic neural network obtained by training in the step (4.1) as new classification networks, keeping the model parameters of the convolutional neural network unchanged in order to improve the network training efficiency, and independently training the model parameters of the deep reinforcement learning cyclic neural network by using the same training set as that in the step (4.1), wherein the training process adopts a reinforcement learning training mode. In order to improve the adaptability of the network to samples, a classification result judgment mode based on reward and punishment rules is introduced: if the identification result belongs to the current multi-component radar signal label set, feeding back a positive reward value to the network, otherwise feeding back a negative penalty value to the network, and only considering the correctness of the final multi-component signal judgment output result in the judgment process, and not considering the output sequence of each signal component judgment result; in order to avoid the overfitting problem which may occur to the network and improve the training efficiency of the network, a memory playback training mode of reinforcement learning is introduced: and storing the state vector input, the classification recognition determination result and the reward and punishment value in a replay memory bank with a certain memory space each time, covering first data in the memory bank by data generated next time after the replay memory bank is full of space, randomly extracting the data in the memory bank in the training process to learn, and disturbing the correlation among the training data. And finally, all parameters of the convolutional neural network and the deep reinforcement learning cyclic neural network are reserved.
(4.3) inputting the samples of the test set into a classifier consisting of a convolutional neural network and a deep reinforcement learning cyclic neural network one by one, outputting a signal type set contained in the current sample by the classifier for many times in a circulating mode, comparing the sample with a standard signal type set corresponding to the sample, and finally obtaining the overall correct recognition probability and the recognition accuracy and the recall rate of each signal type through statistical calculation.
And (4.4) perfecting the network structure and parameters according to the recognition effect test result of the step (4.3), training a new network structure, repeating the network adjustment and training process until the recognition effect test result of the network reaches the expectation, and finishing the training.
And 5: the trained network can realize the intra-pulse modulation mode identification of the multi-component radar signal time-frequency image, single-component or overlapped multi-component radar signals are randomly generated, the time-frequency image subjected to time-frequency analysis and time-frequency image preprocessing is used as the input of the network, and finally the network outputs a signal type set contained in the current radar signal through multiple cycle iteration, so that the intra-pulse modulation mode identification of the multi-component radar signal is realized.
Specifically, in the present embodiment, verification is performed by simulation:
the simulation radar modulation signals are totally 8, the types and the parameters are shown in table 1, and the number of signal sampling points is N = 1024-2048. The signal-to-noise ratio of the training set samples ranges from-6 dB to 10dB, 4000 samples of the single-component signal or the random overlapping multi-component signal with parameters meeting the requirements of Table 1 are generated every 2dB, wherein the sample proportion of the single-component signal to the overlapping multi-component signal is 1. The signal-to-noise ratio range of the test set samples is-10 dB to 10dB, 44000 single-component signal samples and overlapped multi-component signal samples are respectively generated in the same mode as the training set, and 88000 samples are generated in total to serve as the test set.
TABLE 1 simulation Radar Signal parameter Table
Further, fig. 4 to 7 show the recognition effect of the trained neural network under different signal-to-noise ratios in the embodiment of the present invention. When the signal-to-noise ratio is-6 dB, the average correct identification rate of the overlapped multi-component radar signals reaches 94.13%, and the identification accuracy rate and the recall rate of each radar signal type reach more than 90%; and for the single-component signal, when the signal-to-noise ratio is-6 dB, the average correct identification rate of the single-component radar signal reaches 96.30 percent.
The method is effective, can realize the intra-pulse modulation mode identification of 8 types of random overlapping multi-component radar signals under the condition of low signal to noise ratio, and has adaptability to the intra-pulse modulation mode identification of single-component radar signals.
Other step details and functions of the identification algorithm of the intra-pulse modulation mode of the multi-component radar signal in the embodiment of the present invention are known to those skilled in the art, and are not described herein in detail in order to reduce redundancy.
The above description is only an embodiment of the present invention, and the scope of the present invention is not limited thereto, and any person skilled in the art should modify or replace the present invention within the technical specifications of the present invention.
In conclusion, the invention provides a multi-component radar signal intra-pulse modulation mode identification algorithm based on a convolutional neural network and reinforcement learning. Firstly, obtaining a time-frequency image of a radar signal by utilizing Cohen type time-frequency distribution; secondly, suppressing time-frequency image noise by using two-dimensional wiener filtering, adjusting the size and the amplitude of the time-frequency image, adding a label to the time-frequency image, and manufacturing a data set; designing a multi-component radar signal time-frequency image recognition network, pre-training a convolutional neural network by using a training set, extracting radar signal time-frequency image characteristics, further training a multi-component signal classification network by using reinforcement learning, testing the recognition effect of the network by using a test set, and perfecting a network structure and parameters according to a test result; and (3) inputting the radar signal to be identified into a trained multi-component radar signal time-frequency image identification network after time-frequency analysis and image preprocessing, and automatically outputting the types of signal components contained in the radar signal by the network to finish identification. The multi-component radar signal identification algorithm has a wide radar signal type adaptation range and a high identification accuracy under the condition of low signal-to-noise ratio, realizes the identification of the intra-pulse modulation mode of randomly overlapped multi-component radar signals, and has adaptability to the identification of the intra-pulse modulation mode of single-component radar signals.
Claims (3)
1. A method for identifying an intra-pulse modulation mode of a multi-component radar signal is characterized by comprising the following steps:
the method comprises the following steps: acquiring time-frequency images of single-component or overlapped multi-component radar signals of different intra-pulse modulation modes, wherein the time-frequency images comprise LFM signals, MP signals, SFM signals, BPSK signals, 2FSK signals, 4FSK signals, EQFM signals and Frank signals, the signals are used as sample signals, and signals received by a radar receiver are converted by using time-frequency distribution to obtain radar signal time-frequency images;
the specific method of the time-frequency distribution comprises the following steps:
for a received signal x (t), cohen type time frequency distribution is adopted, and the mathematical expression is as follows:
in the above formula, t and ω represent time-frequency distribution independent variable time and angular frequency, Φ (τ, v) is called kernel function, and according to distribution characteristics of radar signals and cross terms in a fuzzy domain, for signals with different frequency modulation slopes, it is proposed to use two kernel functions to respectively obtain time-frequency distribution of radar signals, and the expression of the kernel function is as follows:
in the above formula, α, β, γ and e are used to adjust the size of the kernel function, and the Cohen-type time-frequency distribution images of the radar signal under the two kernel functions are respectively obtained and are jointly used as the time-frequency distribution information of the current radar signal;
step two: preprocessing the radar signal time-frequency image obtained in the first step by using an image processing algorithm, inhibiting noise of the time-frequency image through two-dimensional wiener filtering, adjusting the size of the time-frequency image and normalizing the amplitude, taking the preprocessed radar signal time-frequency image as a sample, taking the signal type contained in the radar signal as a label, and making a training set and a test set;
step three: according to algorithm requirements, the overall network extracts time-frequency image features and carries out multi-label classification recognition according to the extracted features, a pre-training network based on a convolutional neural network is designed to extract radar signal time-frequency image features, and a multi-component signal classification network based on reinforcement learning is designed to obtain classification recognition results;
step four: training the network architecture designed in the third step according to the training set obtained in the second step, testing the network identification effect by using the test set obtained in the second step, and perfecting the network structure and parameters according to the test result to obtain a trained network;
step five: and (4) processing any random overlapping signal in the first step by the second step to be used as a radar signal time-frequency image sample to be identified, inputting the radar signal time-frequency image sample into the network trained in the fourth step, wherein the trained network can provide a radar signal type set contained in the current signal according to the input, and thus, the classification and identification of the multi-component signals are realized.
2. The method for identifying the intra-pulse modulation mode of the multi-component radar signal as claimed in claim 1, wherein: in the third step, the specific method for designing the network architecture is as follows:
a) Designing a convolutional neural network architecture for feature extraction, wherein the convolutional neural network architecture consists of 3 convolutional layers and 3 pooling layers, and vectorizing the output of the last layer of the network to obtain a feature vector of the current time-frequency image;
b) Taking a structure of a full connection layer and a Softmax layer connected behind the full connection layer as a classification network unit, constructing a plurality of classification network units to respectively correspond to label outputs of multi-component signals, classifying the characteristic vectors obtained in the step a), and taking the characteristic vectors and a convolutional neural network framework together as a pre-training time-frequency image characteristic extraction network;
c) Designing a deep reinforcement learning cyclic neural network for multi-component signal classification and identification, replacing a classification network unit in the step b), taking the feature vector obtained in the step a) and the classification historical record of the current cyclic step as input together, and outputting a multi-component signal identification result through a multi-cycle iterative classification process.
3. The method for identifying the intra-pulse modulation mode of the multi-component radar signal as claimed in claim 1, wherein: in the fourth step, the specific method for training and testing is as follows:
d) After the second step, the radar signal time-frequency image subjected to image preprocessing is used as the input of a convolutional neural network, each classification network unit outputs a classification label corresponding to one radar signal component, the time-frequency image characteristic extraction network is pre-trained, and the parameters of the convolutional neural network are reserved after training;
e) Secondly, the convolutional neural network and the deep reinforcement learning cyclic neural network obtained by training in the step d) are used as new classification networks, in order to improve the network training efficiency and keep the model parameters of the convolutional neural network unchanged, the model parameters of the deep reinforcement learning cyclic neural network are trained independently by using the training set same as that in the step d), and all the parameters of the final convolutional neural network and the deep reinforcement learning cyclic neural network are reserved;
f) The samples of the test set are input into a classifier consisting of a convolutional neural network and a deep reinforcement learning cyclic neural network one by one, the classifier can output a signal type set contained in the current sample for many times in a circulating manner, the sample is compared with a standard signal type set, and finally the integral correct recognition probability and the recognition accuracy and the recall rate of each signal type are obtained through statistical calculation;
g) Perfecting the network structure and parameters according to the recognition effect test result of the step f), training a new network structure, repeating the network adjustment and training process until the recognition effect test result of the network reaches the expectation, and finishing the training.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910787759.1A CN110532932B (en) | 2019-08-26 | 2019-08-26 | Method for identifying multi-component radar signal intra-pulse modulation mode |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910787759.1A CN110532932B (en) | 2019-08-26 | 2019-08-26 | Method for identifying multi-component radar signal intra-pulse modulation mode |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110532932A CN110532932A (en) | 2019-12-03 |
| CN110532932B true CN110532932B (en) | 2022-12-13 |
Family
ID=68662779
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910787759.1A Active CN110532932B (en) | 2019-08-26 | 2019-08-26 | Method for identifying multi-component radar signal intra-pulse modulation mode |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110532932B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112103971B (en) * | 2020-09-01 | 2023-07-28 | 广西大学 | Vector reinforcement learning control method of power grid frequency modulation type flywheel energy storage system |
| CN112232120B (en) * | 2020-09-10 | 2022-10-28 | 中国人民解放军海军工程大学 | Radar radiation source signal classification system and method based on software radio |
| CN112491677B (en) * | 2020-11-06 | 2022-10-14 | 厦门大学 | CAN bus identification method and device based on physical layer characteristic fingerprint |
| CN112329710A (en) * | 2020-11-25 | 2021-02-05 | 北京邮电大学 | Signal-to-noise ratio evaluation method and device based on unmanned aerial vehicle remote control signal |
| CN113312996B (en) * | 2021-05-19 | 2023-04-18 | 哈尔滨工程大学 | Detection and identification method for aliasing short-wave communication signals |
| CN113376586A (en) * | 2021-06-03 | 2021-09-10 | 哈尔滨工程大学 | Method for constructing classification model of double-component radar signals |
| CN113962260A (en) * | 2021-10-21 | 2022-01-21 | 中国人民解放军空军航空大学 | Radar signal intelligent sorting method based on denoising depth residual error network |
| CN114021603B (en) * | 2021-10-25 | 2024-05-31 | 哈尔滨工程大学 | Radar signal modulation mode identification method based on model re-parameterization |
| CN115712867A (en) * | 2022-11-03 | 2023-02-24 | 哈尔滨工程大学 | Multi-component radar signal modulation identification method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108388834A (en) * | 2017-01-24 | 2018-08-10 | 福特全球技术公司 | The object detection mapped using Recognition with Recurrent Neural Network and cascade nature |
| CN108805039A (en) * | 2018-04-17 | 2018-11-13 | 哈尔滨工程大学 | The Modulation Identification method of combination entropy and pre-training CNN extraction time-frequency image features |
| CN109086700A (en) * | 2018-07-20 | 2018-12-25 | 杭州电子科技大学 | Radar range profile's target identification method based on depth convolutional neural networks |
| CN109726682A (en) * | 2018-12-29 | 2019-05-07 | 南京信息工程大学 | A kind of human motion recognition method towards weak label sensor data |
| CN109766791A (en) * | 2018-12-25 | 2019-05-17 | 哈尔滨工程大学 | A kind of Modulation Recognition of Communication Signal method based on self-encoding encoder |
| US10304208B1 (en) * | 2018-02-12 | 2019-05-28 | Avodah Labs, Inc. | Automated gesture identification using neural networks |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190050729A1 (en) * | 2018-03-26 | 2019-02-14 | Intel Corporation | Deep learning solutions for safe, legal, and/or efficient autonomous driving |
-
2019
- 2019-08-26 CN CN201910787759.1A patent/CN110532932B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108388834A (en) * | 2017-01-24 | 2018-08-10 | 福特全球技术公司 | The object detection mapped using Recognition with Recurrent Neural Network and cascade nature |
| US10304208B1 (en) * | 2018-02-12 | 2019-05-28 | Avodah Labs, Inc. | Automated gesture identification using neural networks |
| CN108805039A (en) * | 2018-04-17 | 2018-11-13 | 哈尔滨工程大学 | The Modulation Identification method of combination entropy and pre-training CNN extraction time-frequency image features |
| CN109086700A (en) * | 2018-07-20 | 2018-12-25 | 杭州电子科技大学 | Radar range profile's target identification method based on depth convolutional neural networks |
| CN109766791A (en) * | 2018-12-25 | 2019-05-17 | 哈尔滨工程大学 | A kind of Modulation Recognition of Communication Signal method based on self-encoding encoder |
| CN109726682A (en) * | 2018-12-29 | 2019-05-07 | 南京信息工程大学 | A kind of human motion recognition method towards weak label sensor data |
Non-Patent Citations (4)
| Title |
|---|
| Radar Signal Intra-Pulse Modulation Recognition Based on Convolutional Neural Network and Deep Q-Learning Network;ZHIYU QU 等;《IEEE Access》;20200319;第8卷;49125-49136 * |
| Radar Signal Intra-Pulse Modulation Recognition Based on Convolutional Neural Network;ZHIYU QU 等;《IEEE Access》;20180828;第6卷;43874-43884 * |
| 一种带探索噪音的深度循环Q网络;刘全 等;《计算机学报》;20190311;第42卷;1-17 * |
| 变分模态分解与Cohen核结合的时频分布交叉项抑制方法;岳应娟 等;《科学技术与工程》;20170930;第17卷(第27期);66-70 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110532932A (en) | 2019-12-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110532932B (en) | Method for identifying multi-component radar signal intra-pulse modulation mode | |
| CN113378632B (en) | Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method | |
| CN109214452B (en) | HRRP target identification method based on attention depth bidirectional cyclic neural network | |
| CN110555841B (en) | SAR image change detection method based on self-attention image fusion and DEC | |
| CN113657491A (en) | Neural network design method for signal modulation type recognition | |
| CN114726692B (en) | SERESESESENet-LSTM-based radiation source modulation mode identification method | |
| CN112749633B (en) | Separate and reconstructed individual radiation source identification method | |
| CN112149526B (en) | Lane line detection method and system based on long-distance information fusion | |
| CN114997326B (en) | Airborne radar working mode identification method based on multi-feature fusion | |
| CN115982613A (en) | Signal modulation identification system and method based on improved convolutional neural network | |
| CN114943245A (en) | Automatic modulation recognition method and device based on data enhancement and feature embedding | |
| CN115565019A (en) | Single-channel high-resolution SAR image ground object classification method based on deep self-supervision generation countermeasure | |
| CN114528876A (en) | Small sample signal automatic modulation identification method based on improved prototype network | |
| CN112183300B (en) | AIS radiation source identification method and system based on multi-level sparse representation | |
| CN113902095A (en) | Automatic modulation identification method, device and system for wireless communication | |
| CN112132207A (en) | Target detection neural network construction method based on multi-branch feature mapping | |
| CN113343796B (en) | Knowledge distillation-based radar signal modulation mode identification method | |
| CN110751201A (en) | SAR equipment task failure cause reasoning method based on textural feature transformation | |
| CN113343924B (en) | Modulation signal identification method based on cyclic spectrum characteristics and generation countermeasure network | |
| CN115422977A (en) | Radar radiation source signal identification method based on CNN-BLS network | |
| CN115329821A (en) | Ship noise identification method based on pairing coding network and comparison learning | |
| CN115147727A (en) | Method and system for extracting impervious surface of remote sensing image | |
| CN113359091B (en) | Deep learning-based multi-kernel function aliasing radar radiation source identification method and system | |
| CN113759356B (en) | Radar target HRRP identification method based on angular domain feature optimization | |
| CN117590334A (en) | Radar radiation source signal identification method based on GCNet lightweight network |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |