CN114564982A - Automatic identification method for radar signal modulation type - Google Patents

Automatic identification method for radar signal modulation type Download PDF

Info

Publication number
CN114564982A
CN114564982A CN202210061559.XA CN202210061559A CN114564982A CN 114564982 A CN114564982 A CN 114564982A CN 202210061559 A CN202210061559 A CN 202210061559A CN 114564982 A CN114564982 A CN 114564982A
Authority
CN
China
Prior art keywords
modulation type
radar
signal
attention
modulation
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.)
Granted
Application number
CN202210061559.XA
Other languages
Chinese (zh)
Other versions
CN114564982B (en
Inventor
罗皓
吴麒
王翔
侯波涛
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.)
CETC 10 Research Institute
Original Assignee
CETC 10 Research Institute
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 CETC 10 Research Institute filed Critical CETC 10 Research Institute
Priority to CN202210061559.XA priority Critical patent/CN114564982B/en
Publication of CN114564982A publication Critical patent/CN114564982A/en
Application granted granted Critical
Publication of CN114564982B publication Critical patent/CN114564982B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Signal Processing (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The automatic identification method for the radar signal modulation type is simple and efficient in identification mode and stable in training. The invention is realized by the following technical scheme: adopting a one-dimensional convolutional neural network to construct an end-to-end signal modulation type identification network model, and using a real part, an imaginary part and a phase of a complex signal as signal representation information; sampling complex signal samples by a radar, carrying out normalization processing on the complex signal samples, building a modulation type identification network model consisting of a residual error-attention convolution block, a full connection layer and a classifier, and outputting a predicted modulation type through feature extraction, feature integration and probability mapping; after the modulation type recognition network model is determined, network model parameters such as a loss function, an optimizer and a learning rate are set, the network model is trained and the network parameters are updated through a back propagation algorithm by taking a minimum loss function as a target, a signal modulation mode classifier of offline training-online detection is obtained, and radar signal modulation type recognition is further achieved.

Description

Automatic identification method for radar signal modulation type
Technical Field
The invention relates to a radar signal modulation type identification method based on a residual error-attention convolutional neural network, in particular to a method capable of accurately identifying the modulation type of a given signal.
Background
The identification of radar radiation source signals is realized by processing and analyzing intercepted radar signals and mining information such as radar types, states, functions and the like, thereby mastering the information. The identification of the radar signal modulation type is the core of the identification of a radiation source, and the identification of the radar radiation source signal obtains the specific attribute of the signal by analyzing the characteristics in the pulse of different modulation waveform signals, so that the identification of the radar modulation type is realized. Therefore, the rapid and efficient radar signal modulation recognition plays a crucial role in the countermeasure in the electromagnetic spectrum domain, and directly influences the accuracy of subsequent tasks, such as electronic countermeasures and the like. The purpose of radar modulation type identification is to identify the modulation type of an intercepted radar signal, and to judge the functions and characteristics of radar radiation source equipment, a waveform interference or anti-interference task is formed. Along with the rapid development of radar technology, the number and the types of used radars are more and more, the modulation mode of radar signals is more and more complex, and the electromagnetic environment is also more severe. Under the electronic environment, due to the various signal intra-pulse modulation forms and the large signal-to-noise ratio (SNR) range, the essential characteristics of the signals cannot be extracted through a single characteristic, the modulation mode of the radar signals is more and more diversified, the SNR (SNR) threshold of the radar signals which normally work is lower and lower, and the failure of traditional radar intra-pulse modulation identification algorithms which only aim at a few signals and have poor noise immunity is caused. It is therefore an important and difficult task to accurately identify the radar signal modulation scheme.
Conventional modulation recognition techniques rely on feature engineering and classifier models such as Pulse Description Word (PDW) based recognition methods, artificial feature extraction and classification, and joint method feature extraction and machine learning classification. These conventional methods have been widely used in this field and have achieved a number of achievements. However, due to its deficiencies in the depth representation of features, adapting to multiple types of signals and complex electromagnetic environments is challenging. In recent years, many scholars have studied modulation recognition techniques based on a representation learning paradigm, including Dictionary Learning (DL) and deep learning, which have been widely applied to the fields of cognitive radio, image recognition, and the like. The dictionary learning can realize the basic feature learning based on data and the discrimination of low-dimensional representation. The performance of dictionary learning depends to a large extent on the selection and optimization of various discriminators, such as conventional constraints and discriminative constraints. However, performing the dimension-reduced DR and DL algorithms independently may result in an inability to fully utilize the discriminatory information of the training data. Although dictionary learning methods are various, their depth representation capability is limited. It is well known that deep learning has stronger depth representation capability and classification performance due to its depth mapping mechanism. The recognition algorithm based on the deep network and the time-frequency characteristics obtains excellent performance. However, these methods still rely to some extent on a priori knowledge and feature enhancement processing. Most of the methods are based on the condition that sample data is sufficient, radar signal identification is carried out in a high SNR environment or under the condition that the SNR is fixed and unchanged, and actually, a signal identification part of a radar reconnaissance system has to face the conditions of a small sample environment, low SNR and large SNR variation range. Furthermore, these methods are relatively poor in performance, difficult to efficiently identify multiple types of signals with high similarity, and still suffer from a large degree of confusion at high SNR.
For radar modulation type identification, the traditional method mainly comprises a likelihood ratio decision method and a feature extraction identification method, wherein the likelihood ratio decision method is based on a deduced modulation signal likelihood ratio function and completes modulation type identification by calculating test statistic and comparing an identification threshold. However, the method needs a large amount of prior information and cannot adapt to a complex and variable electromagnetic environment. The feature extraction and identification method is composed of two parts of target feature extraction and modulation mode classification, and is characterized in that local differences of signals of different modulation types are highlighted by utilizing information such as instantaneous features, statistical features, time-frequency features and the like, so that target signals are represented. Specifically, the transient characteristics refer to transient signal information represented by amplitude, frequency, phase, and the like; the statistical characteristics mainly comprise signal cumulant, high-order moment, statistical quantity and the like; the time-frequency characteristics are signal time-frequency matrixes extracted by analysis methods such as short-time Fourier transform, wavelet transform, Wigner distribution and the like. After the characteristics of the characterization signals are extracted, a modulation type classifier is designed in the characteristic space by a characteristic extraction and identification method, and the signal modulation type identification is realized. However, in the wavelet transform method, in extracting the characteristics in the chirp signal, along with the decrease of the SNR, there is a disadvantage of poor signal-to-noise ratio performance, the identification effect of some types of signals is deteriorated, and most of the characteristic space and the characteristic extraction algorithm are designed manually, which brings high calculation cost and also causes the loss of characteristic information, and is not favorable for further improvement of the accuracy of signal identification under the background of diversified modulation modes. With the increasing complexity of radar signal types, the updating speed is gradually increased, and the algorithms depending on artificial feature extraction are difficult to adapt to the requirement of radar signal intra-pulse modulation identification.
Deep learning has developed into the core of artificial intelligence, and has made breakthrough progress in the fields of computer vision, natural language processing, economics and the like, thereby drawing wide attention of the society and the research community. The convolutional neural network relies on strong feature learning ability and is widely applied to the fields of identification and classification. In the aspect of radar signal modulation type identification, a one-step identification method based on a one-dimensional convolutional neural network mostly takes a complex signal amplitude value or a vector spliced by real parts and imaginary parts as single-channel input, and does not fully utilize the characteristics of complex signals, namely real parts, imaginary parts and phase information. In general, in the case of a small sample, problems such as poor convergence and reduction in timeliness increase as the network deepens. In practical applications, the application of deep networks is limited by conditions such as sample capacity and computational resources. In addition, in the deep convolutional neural network training process, the correlation among convolutional layer output characteristics is mostly ignored in the existing radar signal modulation type identification method, the different contribution degrees of different output characteristic graphs to target identification are not considered, and the problem that the gradient disappears along with the increase of the number of network layers is solved.
Most of the existing one-dimensional convolutional neural network radar signal modulation type identification methods based on deep learning use the amplitude value of a complex signal as a one-dimensional vector or splice the real part and the imaginary part of the complex signal into a one-dimensional vector as single-channel input of the one-dimensional convolutional neural network, but do not fully utilize the characteristics of the complex signal, namely: real, imaginary and phase information. In addition, due to the insufficient depth representation capability, the depth of the network is large, and the deep feature extraction process is time-consuming and is not suitable for the classification task of a low SNR environment and various signal types. In addition, small networks have a greater demand for feature enhancement processing.
Disclosure of Invention
The invention aims to provide an automatic identification method of radar signal modulation types, which has the advantages of simple and efficient identification mode, stable training and high identification accuracy, aiming at the defects in the prior art.
The purpose of the invention is realized by the following technical scheme: an automatic identification method for radar signal modulation types has the following technical characteristics: the modulation recognizer based on the residual error-attention convolutional neural network directly uses radar intermediate frequency signals as input of the neural network, adopts a one-dimensional convolutional neural network to construct an end-to-end signal modulation type recognition network model, collects radar sampling complex signal samples of different modulation modes, divides the whole input set signal sample into a training set, a verification set and a test set, and determines a neural network architecture; the modulation type identification network model is used for carrying out normalization processing on an input radar complex signal set, extracting a real part, an imaginary part and a phase of a radar complex signal to serve as signal representation information, and inputting the representation information into three channels of a neural network; the modulation type identification network model uses a plurality of residual error-attention convolution blocks with one-dimensional deep convolutional neural networks, attention mechanisms and residual error structures, extracts the signal characteristics of each channel through the one-dimensional deep convolutional neural networks, introduces the residual error structures and the attention mechanisms to adjust the characteristic channels, adds the input signals of the first layer of one-dimensional convolutional neural networks in the residual error-attention convolution blocks to the output of the last layer of one-dimensional convolutional neural networks, and focuses the network on more useful characteristic information by utilizing the interaction of a characteristic diagram and channel statistics; secondly, further extracting interested key features through a full-connection layer, integrating output features of a residual error-attention convolution block, obtaining high-layer feature vectors through feature extraction, feature integration and nonlinear mapping, identifying a modulation mode by using a Softmax classifier, calculating the output probability of each feature vector, and outputting a predicted modulation type; after the modulation type recognition network model determines a neural network architecture, network model parameters such as a loss function, an optimizer and a learning rate are set, the neural network parameters are updated and trained by taking a minimum loss function as a target through a back propagation algorithm, the model with the optimal performance is stored, a modulation mode classifier of off-line training-on-line detection is obtained, the radar signal modulation type is automatically recognized, and the radar signal modulation type recognition is further realized.
The invention has the beneficial effects that:
the identification mode is simple and efficient. The method is based on a residual error-attention convolutional neural network, adopts a one-dimensional convolutional neural network to construct an end-to-end signal modulation type, utilizes radar intermediate frequency signals to be directly used as input of the neural network, collects radar sampling complex signal samples of different modulation modes, divides the whole input set signal sample into a training set, a verification set and a test set, normalizes the complex signal samples, inputs real part, imaginary part and phase information of the radar sampling signals into three channels of the neural network, reduces extracted feature vectors, and improves the classification capability and timeliness of small samples. The intermediate frequency signal of the radar is directly used as the input of the neural network, extra Fourier transform is not needed to be carried out on the received data of the radar, manual design and characteristic selection are not needed, and accidental errors caused by artificial subjective factors are avoided. Compared with the method utilizing the two-dimensional convolutional neural network and based on time-frequency spectrum identification, the calculation amount can be reduced. In addition, the training parameters of the one-dimensional convolutional neural network are far smaller than those of the two-dimensional convolutional neural network, so that the calculation amount required by training is further reduced, and the training difficulty of the model is reduced.
The training is stable. In order to better utilize the inherent characteristics of the complex signals, the real part, the imaginary part and the phase information of the radar sampling signals are used as signal characteristic input of three channels, and the data characteristic of each channel is fully learned by extracting the real part, the imaginary part and the phase of the one-dimensional complex signals as three channels of the input of the one-dimensional convolution neural network, so that the characteristic utilization rate of the complex signals is improved. Along with the gradual and stable increase of the training size, the method has a better characteristic extraction effect. The signal characteristics of each channel are extracted through a one-dimensional convolutional neural network, the key characteristics of the channel are highlighted by using an attention mechanism, and the strong adaptability to small samples and low SNR environments is displayed. A residual error structure and an attention mechanism are introduced into a volume block of the neural network model, so that the method is simpler and more efficient, has a better feature extraction effect, improves the network stability through a depth mapping mechanism, and can realize the best performance even if the features are not enhanced. And the network training is more stable. The input of the first layer of one-dimensional convolutional neural network in the convolutional block is added to the output of the last layer of one-dimensional convolutional neural network, the convolutional layer with the extracted features has high timeliness, the problem of gradient explosion or disappearance along with the deepening of the number of network layers is effectively avoided, and the stability of network training is further ensured. In addition, the invention designs a three-channel attention mechanism aiming at the real part, the imaginary part and the phase input of the complex signal, and leads the network to focus on more useful characteristic information through the interaction of the characteristic diagram and the channel statistic, thereby enhancing the characteristic expression capability and improving the discrimination capability of the model to the interested characteristics.
The identification accuracy is high. After the neural network architecture is determined by adopting the modulation type recognition network model, the network parameters are updated and trained stably by taking the minimum loss function as a target through a back propagation algorithm. And (4) finishing automatic signal identification by combining a Softmax classifier, storing a performance optimal model, and finally finishing signal identification by combining the Softmax classifier. The ability to identify multiple types of high similarity signals is significantly better than other methods, especially in low SNR situations. Compared with the existing one-dimensional convolutional neural network radar signal modulation type identification method which mostly takes the amplitude value of the complex signal as a one-dimensional vector or splices the real part and the imaginary part of the complex signal into a one-dimensional vector to be used as single-channel input of the one-dimensional convolutional neural network, the one-dimensional convolutional neural network radar signal modulation type identification method based on deep learning does not fully utilize the characteristics of the real part, the imaginary part and the phase information complex signal, and the classification accuracy is improved. Reduces the need for feature enhancement processing and has strong recognition performance with the support of large samples and computational resources.
The method comprises the steps of collecting radar sampling complex signals of different modulation modes, and dividing a training set, a verification set and a test set; normalizing the acquired complex signals, and then extracting real parts, imaginary parts and phases of the complex signals to be used as three channels input by a neural network; extracting signal characteristics by using a residual error-attention convolution block, integrating the characteristics through a full connection layer, and classifying a modulation mode by using a Softmax classifier; setting network model parameters such as a loss function, an optimizer, a learning rate and the like; additional advantages, objects, and features of training the network, maintaining a performance optimization model will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention.
Drawings
For the purposes of promoting a better understanding of the objects, aspects and advantages of the invention, reference will now be made to the following detailed description taken in conjunction with the accompanying drawings in which:
FIG. 1 is a flow chart of the present invention for automatic identification of radar signal modulation types;
FIG. 2 is a schematic diagram of a modulation type recognition network model according to the present invention;
fig. 3 is a schematic view of the attention mechanism.
The embodiments of the present invention are described below with reference to specific embodiments, and other advantages and effects of the present invention will be easily understood by those skilled in the art from the disclosure of the present specification. The invention is capable of other and different embodiments and of being practiced or of being carried out in various ways, and its several details are capable of modification in various respects, all without departing from the spirit and scope of the present invention. It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present invention in a schematic way, and the features in the following embodiments and examples may be combined with each other without conflict.
Detailed Description
See fig. 1. According to the modulation recognizer based on the residual error-attention convolutional neural network, a radar intermediate frequency signal is directly used as the input of the neural network, an end-to-end signal modulation type recognition network model is constructed by adopting a one-dimensional convolutional neural network, radar sampling complex signal samples of different modulation modes are collected, the whole input set signal sample is divided into a training set, a verification set and a test set, and a neural network architecture is determined; the modulation type identification network model is used for carrying out normalization processing on an input radar complex signal set, extracting a real part, an imaginary part and a phase of a radar complex signal as signal representation information, and inputting the representation information into three channels of a neural network; the modulation type identification network model uses a plurality of residual error-attention convolution blocks with one-dimensional deep convolutional neural networks, attention mechanisms and residual error structures, extracts the signal characteristics of each channel through the one-dimensional deep convolutional neural networks, introduces the residual error structures and the attention mechanisms to adjust the characteristic channels, adds the input signals of the first layer of one-dimensional convolutional neural networks in the residual error-attention convolution blocks to the output of the last layer of one-dimensional convolutional neural networks, and focuses the network on more useful characteristic information by utilizing the interaction of a characteristic diagram and channel statistics; secondly, further extracting interested key features through a full-connection layer, integrating output features of a residual error-attention convolution block, obtaining high-layer feature vectors through feature extraction, feature integration and nonlinear mapping, identifying a modulation mode by using a Softmax classifier, calculating the output probability of each feature vector, and outputting a predicted modulation type; after the modulation type recognition network model determines a neural network architecture, network model parameters such as a loss function, an optimizer and a learning rate are set, the neural network parameters are updated and trained by taking a minimum loss function as a target through a back propagation algorithm, the model with the optimal performance is stored, a modulation mode classifier of off-line training-on-line detection is obtained, the radar signal modulation type is automatically recognized, and the radar signal modulation type recognition is further realized.
In an alternative embodiment, the modulation type identification network model is at a rangeUnder the condition of surrounding signal to noise ratio, radar complex signal samples of various modulation modes are collected in an equivalent manner, the corresponding modulation types are marked, the radar complex signal samples are divided into a training set, a verification set and a test set sample, and the divided training set sample D {(s) { (S)n,yc)}，n∈[1,N]，c∈[1,C]And ensuring that the division ratio of the training set samples to the verification set samples is 8:2, wherein snFor the nth complex signal sample, ycIt indicates that the sample belongs to the C-th modulation mode, N is the number of collected data samples, and C is the number of modulation types.
All received radar complex signals S ═ S of modulation type recognition network model1,s2,…,sN]The method is characterized in that normalization and real part, imaginary part and phase extraction processing are carried out on all radar complex signal samples, normalization processing is carried out on a radar complex signal set S, the real part and the imaginary part of the S are enabled to fall between (0 and 1), network training is facilitated, and the method is specifically represented as follows:
Figure BDA0003478573100000061
wherein the content of the first and second substances,
Figure BDA0003478573100000062
Figure BDA0003478573100000063
representing a complex set, N being the number of samples, T representing the length of each sample,
Figure BDA0003478573100000064
the signal is a normalized radar complex signal. The collection of the data is carried out,
Figure BDA0003478573100000065
representation matrix for representing complex signals
Figure BDA0003478573100000066
Square of two norms.
Radar complex signal set subjected to normalization processing by modulation type identification network model
Figure BDA0003478573100000067
Then, extracting
Figure BDA0003478573100000068
The real part, the imaginary part and the phase form a third-order tensor which simultaneously contains all information of the complex signal
Figure BDA0003478573100000069
Figure BDA00034785731000000610
And is
Figure BDA00034785731000000611
Wherein the content of the first and second substances,
Figure BDA00034785731000000612
representing a set of real numbers. Due to the three-dimensional tensor
Figure BDA00034785731000000613
Simultaneously contains all information of complex signals, and the third-order tensor
Figure BDA00034785731000000614
As the input of the modulation type identifier, the convolutional neural network in the identifier can better extract all information of the complex signal, thereby improving the accuracy of classification.
Collecting and processing modulation type identification network model to obtain signal sample set
Figure BDA00034785731000000615
Then, the nth complex signal is characterized by a matrix
Figure BDA00034785731000000616
Inputting the radar complex signal into a modulation type identifier based on a residual error-attention convolutional neural network, and outputting a modulation type classification result of the radar complex signal through feature extraction, integration and classification.
See fig. 2. The modulation type identifier is composed of a residual error-attention convolution block, a full connection layer and an output layer which selects a Softmax activation functionWherein P residual-attention convolution blocks are shared in the modulation type identifier, and
Figure BDA00034785731000000617
for the input of the pth residual-attention volume block, P ∈ 1, 2, … P. The residual error-attention volume block consists of a one-dimensional convolution neural network, a residual error structure and an attention mechanism module, and specifically, each residual error-attention volume block comprises L layers of one-dimensional convolution neural networks, wherein the mth characteristic output vector of the ith layer of the one-dimensional convolution neural network
Figure BDA00034785731000000618
The concrete formula is as follows:
Figure BDA00034785731000000619
output feature map of layer I
Figure BDA00034785731000000620
Wherein L is (1, 2, … L), M is (1, 2, … M), M represents the number of current layer convolution kernels, sigma (·) is a ReLU activation function,
Figure BDA00034785731000000621
and
Figure BDA00034785731000000622
the first-level trainable weight and the bias are respectively, convolution operator,
Figure BDA00034785731000000623
the K characteristic diagram output by the previous convolutional layer, K is the (1, 2, … K), and K represents the number of the output characteristic diagrams of the previous layer, namely the number of channels of the input characteristic diagram of the current layer. In particular, input to layer 1 of convolutional neural network
Figure BDA0003478573100000071
And is
Figure BDA0003478573100000072
In addition, in order to ensure network connectivity, the output characteristic diagram of each residual error-attention convolution block is kept to be three channels, namely the number of L-th layer convolution kernels of the one-dimensional convolution neural network is 3,
Figure BDA0003478573100000073
then, a residual structure is introduced into the residual-attention convolution block, and the L-th layer output X of the one-dimensional convolution neural network is outputLAnd layer 1 input X0Adding to obtain a feature map Xrest，Xrest＝XL+X0And is and
Figure BDA0003478573100000074
next, key features are highlighted using an attention mechanism in the residual-attention volume block.
See fig. 3. Three channel profile XrestAnd carrying out global average pooling and nonlinear transformation through global maximum pooling, and calculating a characteristic channel statistic z:
z＝δ(f(f(Fmax(Xrest)))+f(f(Fave(Xrest))))，z＝[z1，z2，z3]then, the feature channel statistic z is applied to the feature map to obtain the key feature output by the pth residual error-attention convolution block
Figure BDA0003478573100000075
Figure BDA0003478573100000076
And is
Figure BDA0003478573100000077
Is composed of
Figure BDA0003478573100000078
Then, a feature map of an m channel is output by highlighting key features in a residual error-attention convolution block by using an attention mechanism, then, the modulation type recognition network classifies the features by adopting a Softmax classifier, the output of the neural network is mapped to a probability space, and a vector is output
Figure BDA0003478573100000079
Figure BDA00034785731000000710
The specific calculation formula of (A) is as follows:
Figure BDA00034785731000000711
and is
Figure BDA00034785731000000712
Based on the probability vector output by the neural network, the modulation type modulation identifier can obtain the predicted modulation category
Figure BDA00034785731000000713
Figure BDA00034785731000000714
Wherein δ (-) denotes a sigmoid activation function F (-) denotes a nonlinear transformation function with the activation function ReLU, Fmax(. represents the maximum pooling, Fave(. cndot.) represents the average pooling, and m.cndot.1, 2, … M, j.cndot.1, … C, C represents the number of neurons in the output layer, i.e., the number of modulation classes, WoutAnd boutAre weights and offsets of the output layer, xjAnd e is the base number of the natural logarithm function, and is the function variable of the Softmax function.
After a modulation identifier based on a residual error-attention convolutional neural network determines a network architecture, the classified cross entropy with high calculation efficiency and high convergence speed is adopted as a loss function:
Figure BDA00034785731000000715
where Ω is all trainable parameters of the network and ynAnd
Figure BDA00034785731000000716
respectively representing the true probability distribution and the predicted probability distribution of the nth training sample. In the training process of the modulation recognizer, the set network model parameters comprise:a random gradient optimizer was selected, and it was determined that the training batch was 64, the number of training rounds was 200, the initial value of the learning rate was 0.001, the decay rate was 0.9, and an Adam optimizer was used in this example.
Training start phase, residual-attention convolution block sets data sample D {(s)n,yc) Inputting the training set in the step (b) into a modulation type identification network model, and updating network parameters by using an Adam optimizer by using a back propagation algorithm and taking a minimum loss function as a target in the modulation type identification. And in the training process, the optimal whole modulation type identification network model tested on the verification set is stored. And after the training is finished, inputting the test set into the stored modulation type recognition network model for online detection.
The same or similar reference numerals in the drawings of the above embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there is an orientation or positional relationship indicated by terms such as "upper", "lower", "left", "right", "front", "rear", etc., based on the orientation or positional relationship shown in the drawings, it is only for convenience of description and simplification of description, but it is not an indication or suggestion that the referred device or element must have a specific orientation, be constructed in a specific orientation, and be operated, and therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes, and are not to be construed as limiting the present invention, and the specific meaning of the terms may be understood by those skilled in the art according to specific situations. Wherein the showings are for the purpose of illustrating the invention only and not for the purpose of limiting the same, and in which there is shown by way of illustration only and not in the drawings in which there is no intention to limit the invention thereto; to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of an actual product; it will be understood by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted. And (6) measuring. Finally, the above embodiments are only intended to illustrate the technical solutions of the present invention and not to limit the present invention, and although the present invention has been described in detail with reference to the preferred embodiments, it will be understood by those skilled in the art that modifications or equivalent substitutions may be made on the technical solutions of the present invention without departing from the spirit and scope of the technical solutions, and all of them should be covered by the claims of the present invention.

Claims (10)

1. An automatic identification method for radar signal modulation types has the following technical characteristics: the modulation recognizer based on the residual error-attention convolutional neural network directly uses radar intermediate frequency signals as input of the neural network, adopts a one-dimensional convolutional neural network to construct an end-to-end signal modulation type recognition network model, collects radar sampling complex signal samples of different modulation modes, divides the whole input set signal sample into a training set, a verification set and a test set, and determines a neural network architecture; the modulation type identification network model normalizes an input radar complex signal set, extracts a real part, an imaginary part and a phase of a radar complex signal as signal representation information, and inputs the representation information into three channels of a neural network; the modulation type identification network model uses a plurality of residual error-attention convolution blocks with one-dimensional deep convolutional neural networks, attention mechanisms and residual error structures, extracts the signal characteristics of each channel through the one-dimensional deep convolutional neural networks, introduces the residual error structures and the attention mechanisms to adjust the characteristic channels, adds the input signals of the first layer of one-dimensional convolutional neural networks in the residual error-attention convolution blocks to the output of the last layer of one-dimensional convolutional neural networks, and focuses the network on more useful characteristic information by utilizing the interaction of a characteristic diagram and channel statistics; secondly, further extracting interested key features through a full-connection layer, integrating output features of a residual error-attention convolution block, obtaining high-layer feature vectors through feature extraction, feature integration and nonlinear mapping, identifying a modulation mode by using a Softmax classifier, calculating the output probability of each feature vector, and outputting a predicted modulation type; after the modulation type recognition network model determines a neural network architecture, network model parameters such as a loss function, an optimizer and a learning rate are set, the neural network parameters are updated and trained by taking a minimum loss function as a target through a back propagation algorithm, the model with the optimal performance is stored, a modulation mode classifier of off-line training-on-line detection is obtained, the radar signal modulation type is automatically recognized, and the radar signal modulation type recognition is further realized.
2. The method for automatically identifying a radar signal modulation type according to claim 1, wherein: the method comprises the steps that a modulation type recognition network model collects radar complex signal samples of various modulation modes in an equivalent manner under a certain range of signal-to-noise ratio, marks corresponding modulation types, divides the radar complex signal samples into a training set, a verification set and a test set sample, and divides the divided training set sample into a training set sample D {(s) { (S)n,yc)}，n∈[1,N]，c∈[1,C]And ensuring that the division ratio of the training set samples to the verification set samples is 8:2, wherein snFor the nth complex signal sample, ycIt indicates that the sample belongs to the C-th modulation mode, N is the number of collected data samples, and C is the number of modulation types.
3. The method for automatically identifying a radar signal modulation type according to claim 1, wherein: all received radar complex signals S ═ S of modulation type recognition network model1,s2,…,sN]The method is characterized in that normalization and real part, imaginary part and phase extraction processing are carried out on all radar complex signal samples, normalization processing is carried out on a radar complex signal set S, the real part and the imaginary part of the S are enabled to fall between (0 and 1), network training is facilitated, and the method is specifically represented as follows:
Figure FDA0003478573090000011
wherein the content of the first and second substances,
Figure FDA0003478573090000012
Figure FDA0003478573090000013
representing a complex set, N being the number of samples, T representing the length of each sample,
Figure FDA0003478573090000014
is a normalized radar complex signal set,
Figure FDA0003478573090000015
representation matrix for representing complex signals
Figure FDA0003478573090000021
Square of two norms.
4. A method for automatic identification of a radar signal modulation type according to claim 3, characterized in that: radar complex signal set subjected to normalization processing by modulation type identification network model
Figure FDA0003478573090000022
Then, extracting
Figure FDA0003478573090000023
The real part, imaginary part and phase of (a) form a third order tensor that contains all the information of the complex signal:
Figure FDA0003478573090000024
and is
Figure FDA0003478573090000025
Wherein the content of the first and second substances,
Figure FDA0003478573090000026
representing a set of real numbers.
5. The method for automatically identifying a radar signal modulation type according to claim 4, wherein: collecting and processing modulation type identification network model to obtain signal sample set
Figure FDA0003478573090000027
Then, the nth complex signal is characterized by a matrix
Figure FDA0003478573090000028
Inputting the data into a modulation type identifier based on a residual error-attention convolutional neural network, and extracting featuresIntegrating and classifying, and outputting the modulation type classification result of the radar complex signal.
6. Method for automatic identification of a radar signal modulation type according to claim 1, characterized in that: the modulation type identifier is composed of a residual error-attention convolution block, a full connection layer and an output layer which selects a Softmax activation function, wherein P residual error-attention convolution blocks are shared in the modulation type identifier, and
Figure FDA0003478573090000029
for the input of the pth residual-attention volume block, P ∈ 1, 2, … P.
7. The method for automatically identifying a radar signal modulation type according to claim 1, wherein: the residual error-attention volume block consists of a one-dimensional convolution neural network, a residual error structure and an attention modeling block, and each residual error-attention volume block contains L layers of one-dimensional convolution neural networks, wherein the mth characteristic output vector of the ith layer of the one-dimensional convolution neural network is epsilon (1, 2, … L)
Figure FDA00034785730900000210
Figure FDA00034785730900000211
Output profile of layer I
Figure FDA00034785730900000212
Wherein L belongs to (1, 2, … L), M belongs to (1, 2, … M), M represents the number of convolution kernels of the current layer, sigma (·) is a ReLU activation function,
Figure FDA00034785730900000213
and
Figure FDA00034785730900000214
the first-level trainable weight and the bias are respectively, convolution operator,
Figure FDA00034785730900000215
the K characteristic diagram output by the previous convolutional layer, K is the (1, 2, … K), and K represents the number of the output characteristic diagrams of the previous layer, namely the number of channels of the input characteristic diagram of the current layer.
8. The method for automatically identifying a radar signal modulation type according to claim 1, wherein: input of layer 1 of convolutional neural network
Figure FDA00034785730900000216
And is
Figure FDA00034785730900000217
The output characteristic diagram of each residual error-attention convolution block is kept to be three channels, namely the number of L-th layer convolution kernels of the one-dimensional convolution neural network is 3,
Figure FDA00034785730900000218
then, a residual structure is introduced into the residual-attention convolution block, and the L-th layer output X of the one-dimensional convolution neural network is outputLAnd layer 1 input X0Adding to obtain a feature map Xrest，Xrest＝XL+X0And is and
Figure FDA00034785730900000219
9. the method for automatically identifying a radar signal modulation type according to claim 8, wherein: three channel profile XrestAnd carrying out global average pooling and nonlinear transformation through global maximum pooling, and calculating a characteristic channel statistic z: z is δ (F (F)max(Xrest)))+f(f(Fave(Xrest))))，z＝[z1，z2，z3]Then, the feature channel statistic z is applied to the feature map to obtain the key feature output by the pth residual error-attention convolution block
Figure FDA0003478573090000031
Figure FDA0003478573090000032
And is
Figure FDA0003478573090000033
Is composed of
Figure FDA0003478573090000034
Then, a feature map of an m channel is output by highlighting key features in a residual error-attention convolution block by using an attention mechanism, then, the modulation type recognition network classifies the features by adopting a Softmax classifier, the output of the neural network is mapped to a probability space, and a vector is output
Figure FDA0003478573090000035
Figure FDA0003478573090000036
The specific calculation formula of (A) is as follows:
Figure FDA0003478573090000037
and is provided with
Figure FDA0003478573090000038
Based on the probability vector output by the neural network, the modulation type modulation identifier can obtain the predicted modulation category
Figure FDA0003478573090000039
Figure FDA00034785730900000310
Where δ (-) denotes the sigmoid activation function F (-) denotes a nonlinear transformation function with the activation function ReLU, Fmax(. cndot.) denotes maximum pooling, Fave(. cndot.) denotes mean pooling, mE (1, 2, … M), j ═ 1, … C, C denotes inputNumber of neurons in the layer, i.e. number of modulation classes, WoutAnd boutAs weights and offsets of the output layer, xjAnd e is the base number of the natural logarithm function, and is the function variable of the Softmax function.
10. The method for automatically identifying a radar signal modulation type according to claim 9, wherein: after a modulation identifier based on a residual error-attention convolutional neural network determines a network architecture, the classified cross entropy is adopted as a loss function:
Figure FDA00034785730900000311
where Ω is all trainable parameters of the network and ynAnd
Figure FDA00034785730900000312
respectively representing the true probability distribution and the predicted probability distribution of the nth training sample.
CN202210061559.XA 2022-01-19 2022-01-19 Automatic identification method for radar signal modulation type Active CN114564982B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210061559.XA CN114564982B (en) 2022-01-19 2022-01-19 Automatic identification method for radar signal modulation type

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210061559.XA CN114564982B (en) 2022-01-19 2022-01-19 Automatic identification method for radar signal modulation type

Publications (2)

Publication Number Publication Date
CN114564982A true CN114564982A (en) 2022-05-31
CN114564982B CN114564982B (en) 2023-09-26

Family

ID=81711365

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210061559.XA Active CN114564982B (en) 2022-01-19 2022-01-19 Automatic identification method for radar signal modulation type

Country Status (1)

Country Link
CN (1) CN114564982B (en)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115604061A (en) * 2022-08-30 2023-01-13 电子科技大学（Cn） Radio frequency signal modulation mode identification method based on external attention mechanism
CN115828154A (en) * 2022-11-25 2023-03-21 中山大学 LPI radar signal identification method, system, equipment and storage medium
CN116016070A (en) * 2022-12-05 2023-04-25 天津光电通信技术有限公司 Digital signal blind demodulation system based on category increment learning
CN116047427A (en) * 2023-03-29 2023-05-02 西安电子科技大学 Small sample radar active interference identification method
CN116593971A (en) * 2023-07-13 2023-08-15 南京誉葆科技股份有限公司 Radar signal modulation identification method of instantaneous frequency characteristic
CN117572379A (en) * 2024-01-17 2024-02-20 厦门中为科学仪器有限公司 Radar signal processing method based on CNN-CBAM shrinkage two-class network
CN117671374A (en) * 2023-12-09 2024-03-08 北京无线电测量研究所 Method and device for identifying image target of inverse synthetic aperture radar

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110133599A (en) * 2019-01-08 2019-08-16 西安电子科技大学 Intelligent radar emitter Signals classification method based on long memory models in short-term
CN113050042A (en) * 2021-04-15 2021-06-29 中国人民解放军空军航空大学 Radar signal modulation type identification method based on improved UNet3+ network
US20210396843A1 (en) * 2020-06-18 2021-12-23 Infineon Technologies Ag Parametric CNN for Radar Processing
CN114021459A (en) * 2021-11-05 2022-02-08 西安晟昕科技发展有限公司 Identification method of small sample radar radiation source
CN114118142A (en) * 2021-11-05 2022-03-01 西安晟昕科技发展有限公司 Method for identifying radar intra-pulse modulation type

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110133599A (en) * 2019-01-08 2019-08-16 西安电子科技大学 Intelligent radar emitter Signals classification method based on long memory models in short-term
US20210396843A1 (en) * 2020-06-18 2021-12-23 Infineon Technologies Ag Parametric CNN for Radar Processing
CN113050042A (en) * 2021-04-15 2021-06-29 中国人民解放军空军航空大学 Radar signal modulation type identification method based on improved UNet3+ network
CN114021459A (en) * 2021-11-05 2022-02-08 西安晟昕科技发展有限公司 Identification method of small sample radar radiation source
CN114118142A (en) * 2021-11-05 2022-03-01 西安晟昕科技发展有限公司 Method for identifying radar intra-pulse modulation type

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
SHUNJUN WEI 等: "Self-Attention Bi-LSTM Networks for Radar Signal Modulation Recognition", IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, pages 5160 *
李润东: "\"基于深度学习的通信信号智能盲检测与识别技术研究 \"", 中国博士学位论文全文数据库 (信息科技辑), pages 136 - 45 *
罗皓: "基于深度学习的大规模MIMO毫米波信道估计", 中国优秀硕士学位论文全文数据库 (信息科技辑), pages 136 - 1494 *

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115604061A (en) * 2022-08-30 2023-01-13 电子科技大学（Cn） Radio frequency signal modulation mode identification method based on external attention mechanism
CN115604061B (en) * 2022-08-30 2024-04-09 电子科技大学 Radio frequency signal modulation mode identification method based on external attention mechanism
CN115828154A (en) * 2022-11-25 2023-03-21 中山大学 LPI radar signal identification method, system, equipment and storage medium
CN115828154B (en) * 2022-11-25 2023-10-03 中山大学 LPI radar signal identification method, system, equipment and storage medium
CN116016070A (en) * 2022-12-05 2023-04-25 天津光电通信技术有限公司 Digital signal blind demodulation system based on category increment learning
CN116047427A (en) * 2023-03-29 2023-05-02 西安电子科技大学 Small sample radar active interference identification method
CN116047427B (en) * 2023-03-29 2023-06-23 西安电子科技大学 Small sample radar active interference identification method
CN116593971A (en) * 2023-07-13 2023-08-15 南京誉葆科技股份有限公司 Radar signal modulation identification method of instantaneous frequency characteristic
CN117671374A (en) * 2023-12-09 2024-03-08 北京无线电测量研究所 Method and device for identifying image target of inverse synthetic aperture radar
CN117572379A (en) * 2024-01-17 2024-02-20 厦门中为科学仪器有限公司 Radar signal processing method based on CNN-CBAM shrinkage two-class network
CN117572379B (en) * 2024-01-17 2024-04-12 厦门中为科学仪器有限公司 Radar signal processing method based on CNN-CBAM shrinkage two-class network

Also Published As

Publication number Publication date
CN114564982B (en) 2023-09-26

Similar Documents

Publication Publication Date Title
CN114564982A (en) Automatic identification method for radar signal modulation type
CN113378632B (en) Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method
CN109871875B (en) Building change detection method based on deep learning
CN111461037B (en) End-to-end gesture recognition method based on FMCW radar
CN112036239B (en) Radar signal working mode identification method and system based on deep learning network
CN104038792B (en) For the video content analysis method and apparatus of IPTV supervision
CN112434662B (en) Tea leaf scab automatic identification algorithm based on multi-scale convolutional neural network
CN108171119B (en) SAR image change detection method based on residual error network
CN111985533A (en) Incremental underwater sound signal identification method based on multi-scale information fusion
CN111476158A (en) Multi-channel physiological signal somatosensory gesture recognition method based on PSO-PCA-SVM
CN110728694A (en) Long-term visual target tracking method based on continuous learning
CN111881954A (en) Transduction reasoning small sample classification method based on progressive cluster purification network
CN111222847A (en) Open-source community developer recommendation method based on deep learning and unsupervised clustering
CN114692665A (en) Radiation source open set individual identification method based on metric learning
CN114429152A (en) Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption
Nie et al. Adap-EMD: Adaptive EMD for aircraft fine-grained classification in remote sensing
CN114980122A (en) Small sample radio frequency fingerprint intelligent identification system and method
CN107437091A (en) Multilayer limits the positive and negative class change detecting method of SAR image of Boltzmann machine
CN116611025B (en) Multi-mode feature fusion method for pulsar candidate signals
CN116863327B (en) Cross-domain small sample classification method based on cooperative antagonism of double-domain classifier
CN105006231A (en) Distributed large population speaker recognition method based on fuzzy clustering decision tree
CN117078960A (en) Near infrared spectrum analysis method and system based on image feature extraction
CN112350745B (en) Sorting method of frequency hopping communication radio station
CN113033683B (en) Industrial system working condition monitoring method and system based on static and dynamic joint analysis
Wang et al. FCM algorithm and index CS for the signal sorting of radiant points

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