CN112749633A - Separate and reconstructed individual radiation source identification method - Google Patents

Separate and reconstructed individual radiation source identification method Download PDF

Info

Publication number
CN112749633A
CN112749633A CN202011563263.5A CN202011563263A CN112749633A CN 112749633 A CN112749633 A CN 112749633A CN 202011563263 A CN202011563263 A CN 202011563263A CN 112749633 A CN112749633 A CN 112749633A
Authority
CN
China
Prior art keywords
extraction module
individual
module
signal
feature extraction
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
CN202011563263.5A
Other languages
Chinese (zh)
Other versions
CN112749633B (en
Inventor
庄旭
尹可鑫
甘翼
袁鑫
丛迅超
李贵
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.)
Southwest Electronic Technology Institute No 10 Institute of Cetc
Original Assignee
Southwest Electronic Technology Institute No 10 Institute of Cetc
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 Southwest Electronic Technology Institute No 10 Institute of Cetc filed Critical Southwest Electronic Technology Institute No 10 Institute of Cetc
Priority to CN202011563263.5A priority Critical patent/CN112749633B/en
Publication of CN112749633A publication Critical patent/CN112749633A/en
Application granted granted Critical
Publication of CN112749633B publication Critical patent/CN112749633B/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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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
    • G06N3/048—Activation functions
    • 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)
  • Molecular Biology (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Signal Processing (AREA)
  • Image Analysis (AREA)

Abstract

The individual radiation source identification method based on separation and reconstruction disclosed by the invention can effectively improve the individual identification accuracy of the radiation source. The invention is realized by the following technical scheme that an individual radiation source signal is collected to establish a SepNet deep learning model; carrying out supervision training on a radiation source signal training set, separating individual features in radiation source signals, inputting original signals into a common feature extraction module to extract common features, and extracting high-dimensional feature vectors of training data; and a signal reconstruction module is adopted to reconstruct the characteristic data from the individual characteristic extraction module and the content characteristic extraction module, reconstruct a characteristic diagram of the input signal, and calculate the mean square error loss with the original signal. The classification module accurately classifies the signals and judges the radiation source individual to which the signals belong according to the obtained class probability value; and finally, updating the convolution kernel weight by using a joint optimization network model, and measuring the recognition capability of the SepNet deep learning model by using the real number of input mapping between 0 and 1.

Description

Separate and reconstructed individual radiation source identification method
Technical Field
The invention belongs to the technical field of signal identification, and particularly relates to an individual radiation source identification (SEI) method of a radiation source based on feature separation and reconstruction.
Background
With the great increase of aerial radiation source equipment, the judgment on whether the aerial target radiation source works normally or the attributes of the abnormal aerial target needs to be carried out, and the judgment needs to be based on the individual identification of the aerial radiation source. The SEI is also called radiation source fingerprint identification or specific radiation source identification, and the SEI identifies a target individual by extracting one or more modulation features exhibited by a received signal, and because the features have the characteristics of universality, testability, stability, uniqueness and the like, the features extracted from the signal can reflect the difference of radiation source equipment, so that the individual radiation source emitting the signal can be identified. Is a technique for determining the individual source emitting a signal by extracting information (radiation source fingerprint) reflecting the identity of its target using an individual feature extraction technique for known signal types, all of which are of the same type. The basic idea is as follows: the fingerprint is observed from different 'visual angles' such as a time domain, a frequency domain, a modulation domain, a transform domain and the like, and the fingerprint is extracted by applying a plurality of effective mathematical tools according to different characteristics of the fingerprint shown in the different 'visual angles'. The individual identification features of the radiation source have the characteristics of universality, testability, stability, uniqueness and the like, and the features extracted from the signals can reflect the difference of the radiation source equipment, so that the individual radiation source emitting the signals can be identified. The radiation source fingerprint is the inherent characteristic of the hardware of the emitting device, has the characteristics of being unforgeable, difficult to change, unavoidable and the like, and is attached to the emitting signal in an unintentional modulation mode. Due to the fact that the signal samples are difficult to obtain, and the fingerprint characteristics are greatly influenced by the channel environment.
Since birth, the SEI technology has attracted much attention in the fields of spectrum management, network security, cognitive radio, and the like, because of its unique function of identifying a specific originating individual. The open set radiation source individual identification technology is a key technology for realizing individual identification of a communication radiation source under actual communication conditions. The extraction of individual subtle features of radiation sources of the same type is one of key technologies and is also a difficulty to be solved.
The radiation source feature extraction methods are various, individual nonlinear differences among different radiation source systems are the root cause of signal fingerprint feature generation, and extraction of fine radiation source individual difference features is very difficult. The output frequencies of different frequency sources are different from the standard frequency to a certain extent, or are higher or lower, but not identical. Furthermore, even the same radiation source signal may be distorted to different degrees due to the different pulse amplitudes of different radiation source signals. Under the condition of low signal-to-noise ratio, individual characteristics of communication radiation sources in a steady-state signal are easily covered, so that the individual characteristics are difficult to extract and identify. The unintentional modulation characteristics in the signal steady state are the result of the comprehensive action of various factors, the extraction of the characteristic vector reflecting the individual radiation source is difficult, and the random modulation of the signal steady state further increases the extraction difficulty of the individual characteristic vector, so that the information loss can be caused in the characteristic extraction process. The individual identification accuracy of these methods therefore depends on precise physical modeling and empirical parameters.
The traditional characteristic extraction method provides an individual identification method based on wavelet transformation characteristic extraction for solving the individual identification problem of a communication radiation source under the non-cooperative communication condition. The method based on the fast Fourier transform is to a great extent to identify different individual radiation sources by utilizing the existing expert experience and utilizing the signal characteristics extracted manually. The quality of the artificially extracted features is greatly influenced by artificial subjective factors, and the method has obvious scene limitation and cannot be well popularized in most cases. Therefore, the identification method cannot be applied to other scenes with very different environments and has low accuracy. In addition, extracting a new set of features for an individual identification task of a particular usage scenario is often time consuming and laborious, and therefore this approach has no future potential for development.
In the aspect of feature extraction, features are generally extracted from the transient process of a radiation source, but the transient signal duration is too short, the similarity with noise is high, and great difficulty is brought to signal acquisition and identification under the uncooperative condition. And the steady-state signal has long duration and is easier to intercept and monitor, so that the method for extracting the required characteristics from the steady-state signal has more practical and wide application value. It is not easy to extract the fingerprint characteristics of the transmitter hardware from the steady state signal, especially the modulated signal. The unknown modulation information often masks the fingerprint information, and the external environment (such as severe noise of a channel, multipath fading, etc.) may distort the fingerprint information, which further increases the difficulty of extracting the individual features. Conventional radiation source identification techniques have been unable to accurately classify and identify radiation sources.
In recent years, deep learning has been a remarkable advance in various application fields, particularly, in computer vision and natural language processing. With this successful incentive, much research has focused in recent years on the use of deep learning to address the radiation source individual identification task. The learning process of deep learning is a process of repeatedly adjusting model parameters, and the performance of hardware such as a GPU is improved, so that complex deep learning training becomes possible. The deep learning extracts the optimal characteristics through forward calculation and backward propagation and continuously adjusting parameters so as to achieve the purpose of prediction. Almost all of the effort in deep learning training is to solve for w and b in the neural network. In deep learning, the learning rate is too small, the iteration times are increased, and the training time is prolonged. However, the learning rate is too large, so that the local optimal point is easy to be crossed, and the accuracy is reduced. The training speed is slow, and the root is that the training speed is caused by excessive parameters of the network structure. The deep learning model training is time-consuming, various pain points such as gradient dispersion and overfitting exist in the model training, the convergence rate is too low, the training time is too long, on one hand, the iteration times in the same total training time are reduced, so that the accuracy is influenced, on the other hand, the training times are reduced, and the chance of trying different hyper-parameters is reduced. Therefore, increasing the rate of convergence is a large pain point. When the neural network is trained, the input distribution of each layer is changed. The learning rate is the same whether the input value is large or small, which is obviously wasteful of efficiency. When the input value is too small, the learning rate cannot be set too large in order to guarantee fine adjustment thereof. The current deep learning-based method has two defects. First, all proposed models are shallow in depth, i.e., the model depth typically does not exceed 10 layers. Most of models only comprise two to three feature extraction layers, and the shallow network structure design cannot extract rich, abstract and high-level complex features for emitter identification. Secondly, the individual features of the emitters of the same kind are slightly different, and the current model does not add any proprietary structure to extract the slight features. The existing deep learning approach solutions use architectures that are too simple and prove inefficient in other research areas such as computer vision and natural language processing.
Disclosure of Invention
Aiming at the problem of low individual identification accuracy of a radiation source in the traditional method, the invention provides the individual radiation source identification method which can realize individual identification of the radiation source, has the advantages of short training time, higher accuracy, simple classification, easy training and higher convergence speed, can effectively improve the individual identification accuracy of the radiation source, has stronger robustness and is based on characteristic separation and reconstruction.
In order to achieve the purpose, the invention adopts the technical scheme that: an individual radiation source identification method for separation and reconstruction, characterized by comprising the following steps:
setting a jump-connected residual block according to a depth residual network DRN, respectively forming a public feature extraction module, an individual feature extraction module and a content feature extraction module by the residual block and a maximum pooling layer, forming a classification module by a full connection layer and a BN layer of a full convolution network structure, and forming a signal reconstruction module by using a plurality of deconvolution layers; collecting individual radiation source signals based on a deep convolutional neural network and a residual block in jump connection to mark the category of the radiation source signals, establishing a characteristic diagram of a convolutional layer and a deconvolution layer characteristic diagram corresponding to the characteristic diagram in a mirror image relationship based on collected and stored data and an individual radiation source identification task, performing jump connection, directly adding corresponding pixels, passing through a nonlinear activation layer, and then transmitting into a next-layer SepNet deep learning model; the SepNet deep learning model carries out supervision training on a radiation source signal training set, individual features in radiation source signals are accurately separated by using a common feature extraction module, a content feature extraction module and an individual feature extraction module, original signals are input into the common feature extraction module to extract common features of the original signals, obtained results are respectively sent into the individual feature extraction module and the content feature extraction module, and high-dimensional feature vectors of training data are extracted; the signal reconstruction module is connected with the middle layers of the individual feature extraction module and the content feature extraction module, reconstructs feature data from the individual feature extraction module and the content feature extraction module, reconstructs a feature map of an input signal, calculates the mean square error loss of the input signal and an original signal, calculates the loss of the input signal and a real label by using a cross entropy loss function, inputs the feature map output by the individual feature extraction module into the classification module to accurately classify the signal, judges the category of the signal, extracts the information dimension of the instantaneous phase of the signal as a classification feature, inputs the radiation source signal training set data into an optimized network model, verifies the classification effect of the model on the radiation source signal, and judges a radiation source individual to which the signal belongs according to the obtained category probability value; and finally, a mean square error loss function and a cross entropy loss function are used for jointly optimizing the network model, when each small batch of gradient descent training is performed, the convolution kernel weight is updated by using backward propagation, and a deconvolution decoder uses a SoftMax function to map the input into real numbers between 0 and 1 and ensure that the sum is 1, so that the recognition capability of the SepNet deep learning model is measured.
Compared with the prior art, the invention has the following beneficial effects:
the training time is short. The invention establishes a SepNet deep learning model aiming at an individual radiation source identification task; carrying out supervision training on a SepNet deep learning model according to a radiation source signal training set, establishing the SepNet deep learning model, setting a jump-connected residual block, respectively forming a public feature extraction module, an individual feature extraction module and a content feature extraction module by the residual block and a maximum pooling layer, forming a classification module by a full connection layer and a BN layer, and forming a signal reconstruction module by a plurality of deconvolution layers; the individual identification of the radiation source can be realized, the extremely strong feature coding capability and reconstruction capability of a deep network are integrated, the interactive information of feature maps of all layers of each module is fully utilized, and the network is not deepened or widened. And the jump connection structure with residual errors is added, so that the structural design idea of the whole network not only accords with the basic principle of individual identification of a radiation source open set, but also the simulation result shows that the identification performance of the method on fine characteristics such as phase noise, frequency drift, harmonic distortion and the like is obviously superior to that of the traditional method, and the method has good noise resistance. The recognition rate can be improved, and the extra time overhead can be reduced. The effectiveness of the method is proved by computer simulation experiments and actually measured data calculation results.
The accuracy is higher. The method adopts a deep SepNet deep learning model to accurately separate individual characteristics in a radiation source signal by utilizing a common characteristic extraction module, a content characteristic extraction module and an individual characteristic extraction module, inputs an original signal into the common characteristic extraction module to extract common characteristics of the original signal, respectively sends obtained results into the individual characteristic extraction module and the content characteristic extraction module to extract high-dimensional characteristic vectors of training data, connects the intermediate layers of the individual characteristic extraction module and the content characteristic extraction module to a signal reconstruction module, reconstructs characteristic data from the individual characteristic extraction module and the content characteristic extraction module by the signal reconstruction module, reconstructs characteristic graphs with different dimensions into input signals, identifies the input signals based on individual radiation sources separated and reconstructed by characteristics, and calculates the mean square error loss of the original signal; the individual identification accuracy of the radiation source can be effectively improved, and the method has strong robustness. By using long and short jump connection, the image classification accuracy is improved, the original input signal can be accurately reconstructed, and the model is trained by utilizing a Softmax and cross entropy loss supervision model. Compared with the traditional deep learning model, the accuracy is higher.
The classification is simple, the characteristic diagram output by the individual characteristic extraction module is input into the classification module to accurately classify the signals, the classification of unknown signals is judged, the information dimension of the instantaneous phase of the signals is extracted to be used as the classification characteristic, the problem that the gradient disappears under the condition that the network layer number is deep can be solved based on the deep convolutional neural network and the jump-connected residual block, meanwhile, the backward propagation of the gradient is facilitated, and the training process is accelerated. By transferring the feature map of the convolutional layer to the deconvolution layer, the decoder is helped to have more image detail information and recover better clean images. The use of both long and short connections allows for minimal losses or the highest accuracy. Compared with the traditional method and other individual radiation source identification methods based on deep learning, the method can greatly improve the accuracy of the model, and enables the algorithm and the model to have application capability in a real environment.
Easy training and high convergence rate. The invention eliminates the noise step by step in the case of full convolution. After each convolutional layer, the noise level is reduced and the details of the image content may be lost. The convolution layer reserves main image content, and the deconvolution layer is used for compensating detail information, so that the image content can be well reserved while a good denoising effect is achieved. On the other hand, the convolution layer gradually reduces the size of the characteristic diagram, and the deconvolution layer gradually increases the size of the characteristic diagram, so that the input and output sizes are ensured to be consistent, and the test efficiency under the condition of limited calculation capacity of the mobile terminal can be ensured. Mapping the input to real numbers between 0 and 1 by using a Softmax function, normalizing the guaranteed sum to be 1, and calculating the loss of the real numbers and the real labels by using a cross entropy loss function; and (3) a network model is jointly optimized by using a mean square error loss function and a cross entropy loss function, the data of the radiation source signal test set are input into the optimized network model, the individual radiation source of the signal is judged according to the obtained class probability value, and the back propagation is applied during each small-batch gradient descent training, so that the training is easy. And updating the convolution kernel weight value so as to measure the identification capability of the SepNet deep learning model. Parameters for the deep learning model are effectively reduced on the premise of ensuring the accuracy, the convergence rate is high, and the training time is greatly shortened. A new solution is provided for landing of individual radiation source identification technology.
Drawings
FIG. 1 is a schematic flow chart of the individual feature identification of the separation and reconstruction radiation source of the present invention.
Fig. 2 is a schematic structural diagram of the SepNet deep learning model.
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described with reference to the accompanying drawings.
Detailed Description
See fig. 1. According to the invention, a jump-connected residual block is set according to a depth residual network, a common feature extraction module, an individual feature extraction module and a content feature extraction module are respectively composed of the residual block and a maximum pooling layer, a classification module is composed of a full-connection layer and a BN layer of a full convolution network structure, and a signal reconstruction module is composed of a plurality of deconvolution layers; collecting individual radiation source signals based on a deep convolutional neural network and a jump-connected residual block to mark the category of the radiation source signals, establishing a feature diagram of a convolutional layer and an anti-convolutional layer feature diagram which is in a mirror image relation with the feature diagram of the convolutional layer for jump connection based on the collected and stored data and an individual radiation source identification task, directly adding corresponding pixels, passing through a nonlinear activation layer and then transmitting the result into a next-layer SepNet deep learning model; the SepNet deep learning model carries out supervision training on a radiation source signal training set, individual features in radiation source signals are accurately separated by using a common feature extraction module, a content feature extraction module and an individual feature extraction module, original signals are input into the common feature extraction module to extract common features of the original signals, obtained results are respectively sent into the individual feature extraction module and the content feature extraction module, and high-dimensional feature vectors of training data are extracted; the signal reconstruction module is connected with the middle layers of the individual feature extraction module and the content feature extraction module, reconstructs feature data from the individual feature extraction module and the content feature extraction module, reconstructs a feature map of an input signal, calculates the mean square error loss of the input signal and an original signal, calculates the loss of the input signal and a real label by using a cross entropy loss function, inputs the feature map output by the individual feature extraction module into the classification module to accurately classify the signal, judges the category of the signal, extracts the information dimension of the instantaneous phase of the signal as the classification feature, inputs the data of a radiation source signal training set into an optimized network model, verifies the classification effect of the model on the radiation source signal, and judges the radiation source individual to which the signal belongs according to the obtained category probability value; and finally, a mean square error loss function and a cross entropy loss function are used for jointly optimizing a network model, when each small batch of gradient descent training is performed, the reverse propagation is used for updating the convolution kernel weight, and a deconvolution decoder uses a SoftMax function to map the input into real numbers between 0 and 1 and ensure that the sum is 1, so that the recognition capability of the SepNet deep learning model is measured.
The residual block comprises four convolutional layers, the output characteristic diagram of each convolutional layer is connected to the ith output of the residual block through jumpthOutput data x of layerl；
xl＝H([xl-1,xl-2,...,x0]) (1)
Such that each shallow layer is directly associated with a deep layer
Wherein x islIs the firstthThe output of the layers, H, is a mapping function, such that each shallow layer is directly associated with a deep layer.
Further, the signal reconstruction module calculates the mean square error loss by using the following formula:
Figure BDA0002860974650000061
where yi is the true value of the ith signal sample, yi' is the predicted value of the model for the ith signal, and n is the number of signals in the batch.
Further, the classification module uses a normalized exponential function Softmax to realize multi-classification in the classification process, the normalized exponential function Softmax calculates the ratio Si of the sums of all element indexes by using the following formula, maps some output neurons to real numbers between (0-1), and normalizes to ensure that the sum is 1, so that the sum of the probabilities of multi-classification is also exactly 1:
Figure RE-GDA0002949471360000062
wherein Si represents the ratio of the index of the current element to the sum of the indices of all elements, ViI represents a category index, and the total number of categories is C.
The normalized exponential function Softmax converts the output values of the multiple classes into relative probabilities.
Further, an important concept in the cross entropy loss function information theory is mainly used for measuring the difference between two probability distributions, and is derived from the divergence of relative entropy (KL), if there are two separate probability distributions p (X) and q (X) for the same random variable X, the difference between the two probability distributions can be measured by using the divergence of KL, and the difference between the two probability distributions can be measured by the signal reconstruction module according to the same random variable X having the two separate probability distributions p (X) and q (X) by using the following divergence calculation formula, and the difference between the two probability distributions can be measured by using the divergence of relative entropy KL:
Figure BDA0002860974650000064
where p (x) represents the true distribution of the samples and q (x) represents the distribution predicted by the model.
Further, the signal reconstruction module disassembles the KL divergence formula into the following form:
Figure BDA0002860974650000071
the former H (p (x)) represents information entropy, the latter is cross entropy, and the cross entropy formula is as follows:
Figure BDA0002860974650000072
further, the common feature extraction module adopts 3 residual blocks each including 2 convolutional layers and a maximum pooling layer in 1 jump connection, performs 2-time down-sampling on input data using the convolutional layers with long and short jump connection full convolutional neural networks, the maximum pooling layer and the convolutional layers for 3 times, extracts common features from the input signals, and sends the processed data to the content feature extraction module and the individual feature extraction module respectively.
Further, the content feature extraction module is responsible for extracting content features in the input signal and is composed of 3 residual blocks with maximum pooling layers. In each residual block, firstly 2 high-dimensional convolutional layers added with jump connection are used, and then a maximum pooling layer is used for down-sampling, wherein the maximum pooling layer down-sampling multiple of the first residual block is 4, and the other 2 down-sampling multiples are 2. Since the core of the SepNet structure design is to separate the content features and individual features from the input signal, it is necessary to ensure that the two features do not contain each other's feature information. Therefore, the output data of the 3 largest pooling layers in the content feature extraction module are simultaneously sent to the signal reconstruction module and are jointly reconstructed with the output results of the individual feature extraction modules, so that the content feature extraction module is guided to be capable of separating out the expected content features.
See fig. 2. The SepNet deep learning model comprises: the signal reconstruction module is connected with the first characteristic separation module and the second characteristic separation module, wherein the first characteristic separation module and the second characteristic separation module comprise an individual characteristic extraction module and a content characteristic extraction module which are connected with the output end of the common characteristic extraction module, and a full connection module which is connected with the individual characteristic extraction module, the signal reconstruction module x1 sends the calculated mean square error loss into the common characteristic extraction module, and the selected Anchor point Anchor1 similar Sample1 is sent into the full connection module through the body characteristic extraction module to calculate the cross entropy loss. The Anchor point Anchor is a reference frame with different sizes and different aspect ratios preset on the image, and is very similar to the window size set by the sliding window method. A typical target detection network may have thousands of anchors,. In practice, anchors are very dense, with multiple anchors all being positive samples corresponding to the same object. The area and proportion of the pooled regions are the anchors one by one. For target detection, each sample of the training data set is labeled, and the size, i.e., length and width, of these boxes is known. These boxes can be classified, and the operation mode of sliding window is used, and the mapping point of the center of the current sliding window in the original pixel space is called Anchor.
The foregoing shows and describes the general principles and broad features of the present invention and advantages thereof. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are given by way of illustration of the principles of the present invention, and that various changes and modifications may be made without departing from the spirit and scope of the invention as defined by the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims (10)

1. An individual radiation source identification method for separation and reconstruction, characterized by comprising the following steps:
setting a jump-connected residual block according to a depth residual network DRN, respectively forming a public feature extraction module, an individual feature extraction module and a content feature extraction module by the residual block and a maximum pooling layer, forming a classification module by a full connection layer and a BN layer of a full convolution network structure, and forming a signal reconstruction module by using a plurality of deconvolution layers; collecting individual radiation source signals based on a deep convolutional neural network and a jump-connected residual block to mark the category of the radiation source signals, establishing a characteristic diagram of a convolutional layer and a deconvolution layer characteristic diagram which is in a mirror image relation with the characteristic diagram of the convolutional layer for jump connection based on collected and stored data and an individual radiation source identification task, directly adding corresponding pixels, passing through a nonlinear activation layer, and then transmitting the mixture into a next SepNet deep learning model; the SepNet deep learning model carries out supervision training on a radiation source signal training set, individual features in radiation source signals are accurately separated by using a common feature extraction module, a content feature extraction module and an individual feature extraction module, original signals are input into the common feature extraction module to extract common features of the original signals, obtained results are respectively sent into the individual feature extraction module and the content feature extraction module, and high-dimensional feature vectors of training data are extracted; the signal reconstruction module is connected with the middle layers of the individual feature extraction module and the content feature extraction module, reconstructs feature data from the individual feature extraction module and the content feature extraction module, reconstructs a feature map of an input signal, calculates the mean square error loss with an original signal, calculates the loss between the original signal and a real label by using a cross entropy loss function, inputs the feature map output by the individual feature extraction module into the classification module to accurately classify the signal, judges the category of the signal, extracts the information dimension of the instantaneous phase of the signal as a classification feature, inputs the radiation source signal training set data into an optimized network model, verifies the classification effect of the model on the radiation source signal, and judges the radiation source individual to which the signal belongs according to the obtained category probability value; and finally, a mean square error loss function and a cross entropy loss function are used for jointly optimizing a network model, when each small batch of gradient descent training is performed, the reverse propagation is used for updating the convolution kernel weight, and a deconvolution decoder uses a SoftMax function to map the input into real numbers between 0 and 1 and ensure that the sum is 1, so that the recognition capability of the SepNet deep learning model is measured.
2. The method of identifying individual radiation sources for isolation and reconstruction of claim 1, wherein: the residual block comprises four convolutional layers, the output characteristic diagram of each convolutional layer is connected to the ith final output of the residual block through jumpingthOutput data x of layerl，
xl＝H([xl-1,xl-2,...,x0]) (1)
Such that each shallow layer is directly associated with a deep layer
Where H is the mapping function.
3. The method of identifying individual radiation sources for isolation and reconstruction of claim 1, wherein: the signal reconstruction module calculates the mean square error loss by adopting the following formula:
Figure RE-FDA0002976643660000011
where yi is the true value of the ith signal sample, yi' is the predicted value of the model for the ith signal, and n is the number of signals in the batch.
4. The method of identifying individual radiation sources for isolation and reconstruction of claim 1, wherein: in the classification process, the classification module adopts a normalization index Softmax to realize multi-classification, the Softmax adopts a formula to calculate the ratio Si of the sum of indexes of all elements, some output neurons are mapped to real numbers between (0-1), and the normalization ensures that the sum is 1, so that the sum of the probabilities of the multi-classification is 1
Figure RE-FDA0002976643660000021
Wherein Si represents the ratio of the index of the current element to the sum of the indexes of all elements, i represents the category index, and ViIs the output of the preceding output unit of the classifier, and C is the total number of classes.
5. The method of identifying individual radiation sources for isolation and reconstruction of claim 1, wherein: the normalized exponential function Softmax converts the output values of the multiple classes into relative probabilities.
6. The method of identifying individual radiation sources for isolation and reconstruction of claim 5, wherein: the signal reconstruction module measures the difference between two probability distributions by adopting a divergence calculation formula according to the same random variable X with two independent probability distributions P (X) and Q (X), and measures the difference between the two probability distributions by using the relative entropy KL divergence:
Figure RE-FDA0002976643660000022
where p (x) represents the true distribution of the samples and q (x) represents the distribution predicted by the model.
7. The method of identifying individual radiation sources for isolation and reconstruction of claim 6, wherein: the signal reconstruction module disassembles the KL divergence formula into the following form:
Figure RE-FDA0002976643660000023
the former H (p (x)) represents information entropy, the latter is cross entropy, and the cross entropy formula is as follows:
Figure RE-FDA0002976643660000024
8. the method of identifying individual radiation sources for isolation and reconstruction of claim 1, wherein: the common feature extraction module adopts 3 residual blocks, each residual block comprises 2 convolution layers and 1 maximum pooling layer in jump connection, performs 2-time down-sampling on input data by using the convolution layers, the maximum pooling layers and the convolution layers 3 times, extracts common features from input signals, and respectively sends the processed data into the content feature extraction module and the individual feature extraction module.
9. The method of identifying individual radiation sources for isolation and reconstruction of claim 1, wherein: the content feature extraction module extracts content features in input signals, 2 high-dimensional convolution layers added with jump connection are used for residual blocks consisting of 3 maximum pooling layers, then the maximum pooling layers are used for down-sampling, wherein the down-sampling multiple of the maximum pooling layer of the first residual block is 4, the down-sampling multiple of the rest 2 residual blocks is 2, the content features and the individual features are separated from the input signals through a SepNet structure, output data of the 3 maximum pooling layers in the content feature extraction module are simultaneously sent into the signal reconstruction module and are jointly reconstructed with output results of the individual feature extraction module, and therefore the content feature extraction module is guided to separate the expected content features.
10. The method of identifying individual radiation sources for isolation and reconstruction of claim 1, wherein: the SepNet deep learning model comprises: the signal reconstruction module is connected with the first characteristic separation module and the second characteristic separation module, wherein the first characteristic separation module and the second characteristic separation module comprise an individual characteristic extraction module and a content characteristic extraction module which are connected with the output end of the common characteristic extraction module, and a full connection module which is connected with the individual characteristic extraction module, the signal reconstruction module x1 sends the calculated mean square error loss into the common characteristic extraction module, and the selected Anchor1 Sample1 is sent into the full connection module to calculate the cross entropy loss through the object detection network based on Anchor of the body characteristic extraction module.
CN202011563263.5A 2020-12-25 2020-12-25 Separate and reconstructed individual radiation source identification method Active CN112749633B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011563263.5A CN112749633B (en) 2020-12-25 2020-12-25 Separate and reconstructed individual radiation source identification method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011563263.5A CN112749633B (en) 2020-12-25 2020-12-25 Separate and reconstructed individual radiation source identification method

Publications (2)

Publication Number Publication Date
CN112749633A true CN112749633A (en) 2021-05-04
CN112749633B CN112749633B (en) 2022-05-17

Family

ID=75646025

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011563263.5A Active CN112749633B (en) 2020-12-25 2020-12-25 Separate and reconstructed individual radiation source identification method

Country Status (1)

Country Link
CN (1) CN112749633B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113343868A (en) * 2021-06-15 2021-09-03 四川九洲电器集团有限责任公司 Radiation source individual identification method and device, terminal and storage medium
CN113705446A (en) * 2021-08-27 2021-11-26 电子科技大学 Open set identification method for individual radiation source
CN113780521A (en) * 2021-08-24 2021-12-10 中国人民解放军93114部队 Radiation source individual identification method based on deep learning

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2006130486A2 (en) * 2005-05-27 2006-12-07 The Regents Of The University Of Michigan Integrative and real-time radiation measurement methods and systems
WO2018020115A1 (en) * 2016-07-28 2018-02-01 Microwave Characterization Center Microwave radiometry imaging system and associated imaging method
CN109307862A (en) * 2018-07-05 2019-02-05 西安电子科技大学 A kind of target radiation source individual discrimination method
CN109684995A (en) * 2018-12-22 2019-04-26 中国人民解放军战略支援部队信息工程大学 Specific Emitter Identification method and device based on depth residual error network
CN110838089A (en) * 2019-09-18 2020-02-25 浙江工业大学 Fast image denoising method based on OctBlock dense block
CN111913156A (en) * 2020-08-05 2020-11-10 电子科技大学 Radar radiation source individual identification method based on deep learning model and feature combination
CN112115924A (en) * 2020-09-27 2020-12-22 西安电子科技大学 Radar radiation source identification method based on one-dimensional CNN and LSTM

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2006130486A2 (en) * 2005-05-27 2006-12-07 The Regents Of The University Of Michigan Integrative and real-time radiation measurement methods and systems
WO2018020115A1 (en) * 2016-07-28 2018-02-01 Microwave Characterization Center Microwave radiometry imaging system and associated imaging method
CN109307862A (en) * 2018-07-05 2019-02-05 西安电子科技大学 A kind of target radiation source individual discrimination method
CN109684995A (en) * 2018-12-22 2019-04-26 中国人民解放军战略支援部队信息工程大学 Specific Emitter Identification method and device based on depth residual error network
CN110838089A (en) * 2019-09-18 2020-02-25 浙江工业大学 Fast image denoising method based on OctBlock dense block
CN111913156A (en) * 2020-08-05 2020-11-10 电子科技大学 Radar radiation source individual identification method based on deep learning model and feature combination
CN112115924A (en) * 2020-09-27 2020-12-22 西安电子科技大学 Radar radiation source identification method based on one-dimensional CNN and LSTM

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
BEINKE, D. 等: "Atom Probe Reconstruction With a Locally Varying Emitter Shape", 《MICROSCOPY AND MICROANALYSIS》 *
JAMES J. SOHN 等: "Analytical Low-Dose CBCT Reconstruction Using Non-local Total Variation Regularization for Image Guided Radiation Therapy", 《FRONTIERS IN ONCOLOGY》 *
T HOFMANN 等: "Dose Reconstruction from PET Images in Carbon Ion Therapy: A Deconvolution Approach Using an Evolutionary Algorithm", 《2017 IEEE NUCLEAR SCIENCE SYMPOSIUM AND MEDICAL IMAGING CONFERENCE》 *
幸涛 等: "基于信号特征深度重构的雷达辐射源识别方法", 《现代信息科技》 *
马丹 等: "改进加速鲁棒特征的平面旋转人脸检测算法", 《计算机工程与设计》 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113343868A (en) * 2021-06-15 2021-09-03 四川九洲电器集团有限责任公司 Radiation source individual identification method and device, terminal and storage medium
CN113780521A (en) * 2021-08-24 2021-12-10 中国人民解放军93114部队 Radiation source individual identification method based on deep learning
CN113780521B (en) * 2021-08-24 2022-03-04 中国人民解放军93114部队 Radiation source individual identification method based on deep learning
CN113705446A (en) * 2021-08-27 2021-11-26 电子科技大学 Open set identification method for individual radiation source
CN113705446B (en) * 2021-08-27 2023-04-07 电子科技大学 Open set identification method for individual radiation source

Also Published As

Publication number Publication date
CN112749633B (en) 2022-05-17

Similar Documents

Publication Publication Date Title
CN113378632B (en) Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method
CN112749633B (en) Separate and reconstructed individual radiation source identification method
CN108696331B (en) Signal reconstruction method based on generation countermeasure network
CN114564982B (en) Automatic identification method for radar signal modulation type
CN114692665B (en) Radiation source open set individual identification method based on metric learning
CN113259288B (en) Underwater sound modulation mode identification method based on feature fusion and lightweight hybrid model
CN114595732B (en) Radar radiation source sorting method based on depth clustering
CN111050315B (en) Wireless transmitter identification method based on multi-core two-way network
CN108960142B (en) Pedestrian re-identification method based on global feature loss function
CN112115821B (en) Multi-signal intelligent modulation mode identification method based on wavelet approximate coefficient entropy
CN111612130B (en) Frequency shift keying communication signal modulation mode identification method
CN112329784A (en) Correlation filtering tracking method based on space-time perception and multimodal response
Huang et al. Radar waveform recognition based on multiple autocorrelation images
CN116471154A (en) Modulation signal identification method based on multi-domain mixed attention
CN111242028A (en) Remote sensing image ground object segmentation method based on U-Net
CN114980122A (en) Small sample radio frequency fingerprint intelligent identification system and method
CN113033345B (en) V2V video face recognition method based on public feature subspace
CN116482618A (en) Radar active interference identification method based on multi-loss characteristic self-calibration network
CN116340846A (en) Aliasing modulation signal identification method for multi-example multi-label learning under weak supervision
CN115661539A (en) Less-sample image identification method embedded with uncertainty information
CN115955375A (en) Modulated signal identification method and system based on CNN-GRU and CA-VGG feature fusion
CN110610152B (en) Multispectral cloud detection method based on discriminative feature learning unsupervised network
Guo-Xi Research on a Convolutional Neural Network Method for Modulation Waveform Classification.
CN114785649B (en) Satellite communication signal identification method based on multiport neural network
CN114187528B (en) Hyperspectral target detection method based on multi-example spatial spectrum information joint extraction

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