CN112613443A - Robustness communication radiation source intelligent identification method based on deep learning - Google Patents

Robustness communication radiation source intelligent identification method based on deep learning Download PDF

Info

Publication number
CN112613443A
CN112613443A CN202011592239.4A CN202011592239A CN112613443A CN 112613443 A CN112613443 A CN 112613443A CN 202011592239 A CN202011592239 A CN 202011592239A CN 112613443 A CN112613443 A CN 112613443A
Authority
CN
China
Prior art keywords
sample data
signal sample
signal
data
radiation source
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011592239.4A
Other languages
Chinese (zh)
Inventor
费泽松
王新尧
李维彪
尹睿锐
曾鸣
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.)
Chongqing Innovation Center of Beijing University of Technology
Original Assignee
Chongqing Innovation Center of Beijing University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chongqing Innovation Center of Beijing University of Technology filed Critical Chongqing Innovation Center of Beijing University of Technology
Priority to CN202011592239.4A priority Critical patent/CN112613443A/en
Publication of CN112613443A publication Critical patent/CN112613443A/en
Pending legal-status Critical Current

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
    • G06F2218/02—Preprocessing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • 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
    • 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
    • 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 Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Signal Processing (AREA)
  • Probability & Statistics with Applications (AREA)
  • Monitoring And Testing Of Transmission In General (AREA)

Abstract

The invention discloses a robustness communication radiation source intelligent identification method based on deep learning, which comprises the following steps: preprocessing signal sample data; performing data enhancement on signal sample data; the method comprises the steps of extracting the characteristics of signal sample data by using a CNN module and an LSTM model, classifying the characteristics of each signal sample data by using a Softmax logistic regression model, identifying abnormal signal sample data according to the confidence coefficient of each signal sample data, and performing K-Means clustering on the characteristics of the abnormal signal sample data. The design scheme of the invention is simple and easy to realize, can aim at the received signals of different carrier frequencies, signal-to-noise ratios and phase distortion degrees, has strong adaptability to the intercepted signals under different communication environments, and can not only identify the history radiation sources with labels, but also carry out blind source classification on abnormal radiation sources without labels.

Description

Robustness communication radiation source intelligent identification method based on deep learning
Technical Field
The invention relates to the field of machine learning, in particular to a robust communication radiation source intelligent identification method based on deep learning.
Background
Research on individual identification of communication radiation sources was carried out later, and in 1995, Choe published the first literature on identification of communication radiation sources, and devices were identified based on transient signals of wireless transmitters. Transient state signals are derived from a short unstable state where a communication radio station is in when the communication radio station is just started or is switched to a working mode, after the transient state process is finished, a radiation source enters a long-time stable state, at the moment, fluctuation of each component of the equipment is small, and the components are superposed on a steady state signal layer, so that it is more difficult to analyze and extract fine features related to the identity of the radiation source in the steady state process, but the steady state signals are easy to obtain and still have important research values. At present, analysis on steady-state characteristics mainly comprises signal carrier frequency and code rate, signal instantaneous parameters, stray characteristic analysis based on high-order spectrum and the like.
In the traditional solution, a plurality of characteristics of the radiation source identity contained in the signal need to be extracted manually and accurately, and a classifier needs to be designed to perform fusion analysis on the characteristics so as to develop identification. The classifiers can be classified into supervised, unsupervised and semi-supervised classifiers according to the number of labeled samples in the collected signal samples. The supervised classifier is suitable for the condition of sufficient labeled samples and mainly comprises a K neighbor classifier, a neural network, a decision tree and the like; the semi-supervised classifier is based on a label-free data optimization model, provides gain for the performance of the model under the condition of less label data, and mainly comprises a generating model, a semi-supervised support vector machine and the like; for the case without any label, the method belongs to the category of blind sorting and is usually processed by dimension reduction and clustering algorithms. However, the conventional classification method excessively depends on the accuracy and the cleanliness of the extracted data features, the robustness of the features is poor, and the identification accuracy is greatly reduced due to the occurrence of interference such as small noise, frequency offset and phase shift of signals. Therefore, how to obtain a robust low-complexity radiation source identification method is crucial to practical system application.
With the gradual maturity and falling of Deep Learning (DL) in many fields such as image processing, Computer Vision (CV), Natural Language Processing (NLP), recommendation systems, speech recognition, network security, etc., in recent years, the deep learning has received more and more attention in the field of wireless communication. The neural network has high nonlinear fitting capability and complex deep feature characterization capability, and can mine data features from different dimensions. The Convolutional Neural Network (CNN) can fully mine the spatial local characteristics of data by using the local receptive field; the Recurrent Neural Network (RNN) and the derivative long-term memory network (LSTM) thereof can utilize a special forgetting gate mechanism to mine the front and back relevant characteristics of a time sequence and discard irrelevant characteristics to improve the training efficiency. Therefore, it may be considered to solve the robust radiation source identification problem with deep neural networks.
Disclosure of Invention
The invention aims to: aiming at the existing problems, the robustness communication radiation source intelligent identification method based on deep learning is provided to accurately identify the communication radiation source and carry out blind sorting on unknown radiation sources.
The technical scheme adopted by the invention is as follows:
a robustness communication radiation source intelligent identification method based on deep learning comprises the following steps: preprocessing signal sample data; performing data enhancement on signal sample data by utilizing the signal-to-noise ratio of the test set sample evaluation acquired in the test environment; performing convolution operation on signal sample data by using a CNN (convolutional neural network) module to extract signal space characteristics, extracting time domain correlation characteristics of the signal sample data by using an LSTM (linear spline) module, splicing the CNN module and the LSTM module on an output dimension, integrating the characteristics of each signal sample data through a full connection layer, and classifying the characteristics of each signal sample data by using a Softmax logistic regression model, wherein the classification process comprises the process of calculating the confidence coefficient of each signal sample data; acquiring abnormal signal sample data according to the relation between the confidence coefficient of each signal sample data and the first condition; and performing K-Means clustering on the characteristics of the abnormal signal sample data.
Further, the preprocessing step of the signal sample data includes: and sequentially carrying out carrier frequency estimation, down-conversion, phase compensation, down-sampling and data standardization processing on the signal sample data.
Further, the data normalization process includes: the mean of the data in each dimension is calculated using the totality of the signal sample data, subtracted in each dimension, and divided by the standard deviation of the data in that dimension.
Further, the CNN module stacks 3 one-dimensional convolution layers, the sizes of convolution kernels are sequentially reduced, the activation functions of the convolution layers use ReLU, batch normalization is carried out before activation, and after the convolution layers are stacked, a one-dimensional global average pooling layer is followed, and output corresponds to the feature channel dimension.
Further, the LSTM module divides the input signal into 4 time steps, the dimension of the hidden state is 128, a Dropout layer with a drop probability of 0.5 is added, and the hidden state of the last time step is used as the output of the LSTM module.
Further, the signal-to-noise ratio of the test set sample collected in the test environment is obtained by evaluating a signal-to-noise ratio estimation method based on a high-order moment.
Further, the data enhancement on the signal sample data includes: white gaussian noise corresponding to the estimated signal-to-noise ratio is added to the signal sample data.
Further, the method for calculating the confidence of each signal sample data comprises the following steps: and taking the reciprocal of the information entropy of the probability distribution output by the Softmax logistic regression model as the confidence coefficient of the corresponding signal sample data.
Further, the clustering number of the K-Means clustering is obtained by calculating based on the sum of squared errors by adopting an elbow method.
Further, the characteristic of the abnormal signal sample data subjected to K-Means clustering is extracted by using the spliced CNN module and LSTM module and integrated by a full connection layer.
In summary, due to the adoption of the technical scheme, the invention has the beneficial effects that:
1. the identification process and the classification model designed by the invention are easy to realize, are suitable for receiving signals with different carrier frequencies, signal-to-noise ratios and phase distortion degrees, and have strong adaptability to intercepted signals in different communication environments.
2. The method can identify the history radiation source with the label and can also carry out blind source classification on the abnormal radiation source without the label.
3. The method of the invention designs a preprocessing step of the sample data, cleans the training data, is beneficial to deep neural network to extract the subtle characteristics of the radiation source, and ensures that the model identification sensitivity is high.
4. The method of the invention designs a sample enhancement step, so that the trained deep learning model can be better adapted to the actual test environment, thereby improving the matching degree of learning and the actual environment.
5. The invention sets the optimal cluster number, so that the identification result is more reasonable, and the identification result of the method is closer to the actual environment.
Drawings
The invention will now be described, by way of example, with reference to the accompanying drawings, in which:
fig. 1 is a schematic diagram of a signal sample data preprocessing flow.
Fig. 2 is a schematic diagram of the digital down conversion principle.
Fig. 3 is a schematic diagram of a signal sample data feature extraction neural network topology.
Detailed Description
All of the features disclosed in this specification, or all of the steps in any method or process so disclosed, may be combined in any combination, except combinations of features and/or steps that are mutually exclusive.
Any feature disclosed in this specification (including any accompanying claims, abstract) may be replaced by alternative features serving equivalent or similar purposes, unless expressly stated otherwise. That is, unless expressly stated otherwise, each feature is only an example of a generic series of equivalent or similar features.
Example one
The embodiment discloses a robustness communication radiation source intelligent identification method based on deep learning, which comprises the following steps:
A. data pre-processing
As shown in fig. 1, in the signal preprocessing stage, the raw data is processed into baseband signal data that is beneficial for the deep neural network to extract the fine features of the radiation source through multiple stages of carrier frequency estimation, down conversion, phase calculation, down sampling, data normalization, and the like.
A1. Carrier frequency estimation
Because of the unknown frequency offset, the carrier frequency of the signal is first estimated before down-converting the intermediate frequency signal. For a QPSK signal, a quadric spectrum of the QPSK signal includes a direct current component, and a strong discrete spectrum component also appears at a quadrupling carrier frequency, and a carrier frequency estimation principle based on the quadric spectrum is simple, has a small computation amount and high accuracy, and can accurately measure a carrier frequency of the signal according to the characteristic.
The fourth power spectrum is defined as
X(k)＝fft(x4(n)) (1)
A2. Down conversion
The down-conversion aims to remove the intermediate frequency carrier of the received signal, and the frequency spectrum of the signal is moved to a baseband by using a digital quadrature down-conversion method according to the carrier frequency value obtained by the fourth power spectrum estimation in the step (1), and simultaneously, the in-phase component and the quadrature component of the signal are extracted as original data. Wherein the intermediate frequency signal can be expressed as
Figure BDA0002867227270000051
The intermediate frequency signal is multiplied by the cosine signal and the sine signal of the same frequency respectively, and low-pass filtering is performed to obtain the in-phase component and the quadrature component of the signal, and the signal is down-converted to the baseband to obtain the quadrature component i (t) and the in-phase component q (t), as shown in fig. 2.
A3. Phase calculation
And (3) calculating a phase value of each sampling point of the I/Q sampling data after down-conversion, and compensating the phase to be used as the characteristic vector input into the neural network in the step (2). For example: for QPSK signals, the phase can be compensated by calculating the average phase angle of each constellation point in the first quadrant of the constellation diagram and comparing its difference with 45 °.
A4. Down sampling
For an original intermediate frequency sampling signal, each modulation symbol comprises a plurality of sampling points, the signal-to-noise ratio can be maximized only by sampling and outputting at the optimal sampling moment of a receiver, and partial interference can be filtered out while the information redundancy of the same modulation symbol is reduced by analyzing and down-sampling the optimal sampling moment of the received signal. The method comprises the steps of calculating the variance of the module of the sampling point of the signal sample at different sampling moments, estimating the optimal sampling moment in the signal sample duration, sorting the variances of the modules of the sampling points at all different sampling moments, and selecting the sampling points according to the sequence of the variances from small to large by combining with the down-sampling ratio.
A5. Data normalization
The data standardization uses all sample data to calculate the mean value of the data on each dimension, then the mean value is subtracted from each dimension, and finally the standard deviation of the data on the dimension is divided by each dimension of the data, so that the features of each dimension conform to the standard normal distribution N (0,1), the features of each dimension are normalized to the same value interval, and the training efficiency and the training stability are improved.
B. Communication radiation source individual recognition model design based on deep learning
The embodiment uses CNN to perform convolution operation on the signal to extract the signal space characteristic, and uses LSTM to extract the time domain correlation characteristic of the communication signal. And the idea of parallel connection network in GoogleNet is combined to realize a radiation source identification model based on CNN and LSTM common design. The number of input phase characteristic sampling points is 1024, and two blocks, CNN-Block and LSTM-Block, are designed in parallel, as shown in FIG. 3.
CNN-Block design
The first module is CNN-Block, 3 one-dimensional convolutional layers (Conv-layer) are stacked, the sizes of convolutional cores are reduced in sequence, ReLU is used as an activation function of the convolutional layers, Batch-normalization (Batch-normalization) is carried out before activation, after the convolutional layers are stacked, a one-dimensional global average pooling layer (Avg-pooling-layer) is followed, output corresponds to the set feature channel dimension, parameters are reduced, and effective features are extracted.
LSTM-Block design
The second module is LSTM-Block, and divides the input signal into 4 Time steps (Time-slots), the dimension of the hidden state is 128, a Dropout layer with a discarding probability of 0.5 is added, and the hidden state of the last Time step is taken as the output of the module.
B3. Module fusion
And splicing the two parallel modules B1 and B2 on the output dimension to realize the fusion of the features, and using Softmax output to identify and judge after the features are integrated through a full connection layer.
C. Robust model generalization under signal-to-noise ratio variation
C1. Signal-to-noise ratio estimation
In order to enable the trained deep learning model to be better adapted to the actual test environment, the signal-to-noise ratio of the received data of the actual test environment is estimated, and data enhancement processing is conveniently carried out on the training data based on the estimation result.
In a non-cooperative communication environment, prior information of a received signal is difficult to obtain, for a constant envelope modulation signal, a signal-to-noise ratio estimation method based on a high-order moment can be used, the method is simple and efficient, the prior information is not required to be relied on, an intermediate frequency sampling signal provided by a data set can be regarded as a narrow-band signal interfered by noise, and if the noise is additive white Gaussian noise, the noise exists
r(t)＝s(t)+n(t) (3)
Wherein
Figure BDA0002867227270000081
Figure BDA0002867227270000082
The second and fourth moments of the intermediate frequency signal can be expressed as
m2＝E[r*(t)r(t)]
m4＝E{[r*(t)r(t)]2} (5)
On the other hand, note PsAnd PnThe power of signal and noise respectively, according to the narrow-band random signal theory, for the modulation signal with constant envelope such as QPSK, there are
m2＝Ps+Pn
Figure BDA0002867227270000083
Combining the above two forms to obtain
Figure BDA0002867227270000084
The carrier frequency of the intermediate frequency signal is offset in the solving process, and the signal-to-noise ratio estimated by the method is not influenced by the carrier frequency estimation deviation. For the real intermediate frequency signal, Hilbert conversion is carried out on the real intermediate frequency signal to obtain a corresponding complex intermediate frequency signal, and the signal-to-noise ratio can be estimated according to the method.
C2. Sample enhancement
Estimating the range of the signal-to-noise ratio of the signal sample of the test set, performing sample enhancement on the training data based on the estimation result, namely adding corresponding white Gaussian noise to the training data to ensure that the signal-to-noise ratio of the signal sample of the training data is consistent with that of the test set, and training a model by using the training data after noise addition.
D. Anomaly detection
The method can also mark a part of special signals which are received by the model and have certain difference with the training data as abnormal signals in the identification process.
D1. Anomaly detection algorithm based on confidence scores
On the basis of identifying the test signal sample by using the model, the model outputs a confidence score value for measuring the degree of confidence of the model on the current identification result, if the confidence score exceeds a confidence threshold value, the signal sample is considered to be normal data, and is identified according to the identification result of the model, otherwise, the signal sample is abnormal data.
The output of the model is probability distribution, the information entropy of the probability distribution reflects the chaos degree of the distribution, the larger the information entropy is, the higher the chaos degree is, the larger the uncertainty is, and therefore the reciprocal of the information entropy of the output distribution can be regarded as a confidence score. The mathematical expression for abnormality detection by this method is as follows
Figure BDA0002867227270000091
Figure BDA0002867227270000092
Wherein x is(i)For the input of the ith test sample, hθRepresenting a trained network model, c (x)(i)) Denotes the confidence score, λ is the confidence threshold, f (x)(i)) Is an exception flag.
D2. Clustering feature selection
Before clustering processing is carried out on abnormal signals, firstly, clustering features are selected, when a deep neural network is trained, features of the signals are extracted by a network model, and 256-dimensional output of the model in a Concat layer can be directly adopted as the clustering features of the signals.
D3. Elbow method for judging optimal cluster number
The elbow method determines the optimal cluster number based on Sum of Squares of Errors (SSE). The method comprises the steps of gradually increasing the value of a cluster number k from 1 cluster, sequentially calculating SSEs when the k takes different values, increasing the k, enabling the aggregation degree of samples of each cluster to be tighter, further enabling the value of the SSEs to be gradually reduced, when the value of the k is still smaller than the real cluster number, enabling the increase of the k to provide larger gain for the aggregation degree of the samples in each cluster, enabling the SSEs to relatively decrease at a higher speed, enabling the contribution of the increase of the k to the aggregation degree of each cluster to be sharply reduced and to slowly and smoothly go on subsequent change once the k is increased to the real cluster number, and enabling a function curve of the SSEs and the cluster number k to be in the shape of an elbow, and determining the optimal cluster number only by observing the k corresponding to the elbow. The computational expression of SSE is
Figure BDA0002867227270000101
Wherein, CiDenotes the ith cluster, p denotes the cluster CiSample point of (1), miIs a cluster CiThe center of mass of the lens.
D4.k-means clustering
After the optimal clustering number is determined, blind sorting can be performed on abnormal signals by using a clustering algorithm, the k-means clustering algorithm can simply and quickly cluster samples according to clustering characteristics on the premise of giving the clustering number, and the method comprises the following steps:
step 1: randomly initializing k points as initial cluster centers. The clustering center is a position coordinate with the same vector length as each data sample;
step 2: calculating the distance from each data sample to each clustering center, and for any data sample, dividing the data sample into which cluster the data sample is closest to which clustering center;
step 3: calculating the coordinate mean value in each cluster as a new cluster center of the cluster;
step 4: the above steps are repeated until the center of each class does not change much after each iteration.
After clustering, blind sorting of the abnormal communication radiation sources can be realized, and thus, through the steps A-D, intelligent identification of the communication radiation source individuals based on deep learning is completed.
Example two
The embodiment discloses a robustness communication radiation source intelligent identification method based on deep learning, which comprises the following steps:
and preprocessing signal sample data. The preprocessing step performs data cleaning on the signal sample data so that the sample data is convenient for training the learning model. The preprocessing step comprises the processes of carrier frequency estimation, down-conversion, phase compensation, down-sampling and data standardization processing.
After preprocessing, signal sample data is subjected to data enhancement by utilizing the signal-to-noise ratio of the test set sample evaluation acquired in the test environment. The test environment is an application environment, and the test set samples are also signal samples collected in the application environment, so that the signal samples serve as data enhancement references of the training samples, and the training samples are more adaptive to the actual environment.
Based on a designed model architecture, a CNN module is used for carrying out convolution operation on signal sample data to extract signal space characteristics, an LSTM module is used for extracting time domain correlation characteristics of the signal sample data, the CNN module and the LSTM module are spliced on an output dimension, the characteristics of the signal sample data are integrated through a full connection layer, a Softmax logistic regression model is used for classifying the characteristics of the signal sample data, and the classification process comprises the process of calculating the confidence coefficient of the signal sample data. Here, the classification of the model may classify known radiation sources.
And acquiring abnormal signal sample data according to the relation between the confidence coefficient of each signal sample data and the first condition. The so-called first condition, which is different from the condition satisfied by the conventional radiation source signal, generally speaking, a higher confidence level indicates that the sample tends to be about a conventional signal, and therefore, the first condition may be set as a threshold, and signal sample data with a confidence level below the set threshold may be determined as abnormal signal sample data.
And performing K-Means clustering on the characteristics of the abnormal signal sample data. The K-Means clustering is provided with the clustering number, and the optimal clustering number is preferably judged by adopting an elbow method. The elbow method determines the optimal cluster number based on Sum of Squares of Errors (SSE). The method comprises the steps of gradually increasing the value of a clustering cluster number k from 1 cluster, sequentially calculating SSE when the k takes different values, increasing the k, enabling the aggregation degree of samples of each cluster to be tighter, further enabling the value of the SSE to be gradually reduced, when the value of the k is still smaller than the real clustering cluster number, enabling the increase of the k to provide larger gain for the aggregation degree of the samples in each cluster, enabling the SSE to relatively decrease at a higher speed, enabling the contribution of the increase of the k to the aggregation degree of each cluster to be sharply reduced and slowly and smoothly trend in subsequent changes once the k is increased to the real clustering cluster number, and enabling a function curve of the SSE and the clustering cluster number k to be in the shape of an elbow, and determining the optimal clustering cluster number only by observing the k corresponding to the elbow of the hand.
EXAMPLE III
The present embodiment explains the identification process by taking data collected by an actual device as an example. Because the data used are acquired by actual equipment, certain noise, frequency offset and phase offset exist.
The embodiment totally contains signal samples generated by 30 communication radiation sources, the signals are all sampled by 70MHz intermediate frequency signals, the sampling rate is 30MSps, the signal carrier frequency is 1000MHz, the modulation mode is QPSK, the modulation rate is 1M Baud/s, and unknown radiation source signals which are not contained in training data are in a blind sorting module in test data.
Firstly, preprocessing the acquired data signals, as shown in fig. 1, the specific training steps of the deep learning network are as follows:
step one, carrier frequency estimation
For QPSK signals, a strong discrete spectral component appears at quadruple carrier frequencies, from which the carrier frequency of the signal is accurately measured.
The fourth power spectrum is defined as
X(k)＝fft(x4(n)) (1)
Step two, down conversion
As shown in fig. 2, the coherent detection method is used to multiply the intermediate frequency signal by the cosine signal and the sine signal of the same frequency, and low-pass filter the resulting signal to obtain the in-phase component and the quadrature component of the signal, and down-convert the signal to the baseband.
Step three, phase calculation
And C, calculating the phase value of each sampling point for the I/Q sampling data after the lower edge frequency, and compensating the phase to be used as the characteristic vector input into the neural network in the step six. For QPSK signals, the phase can be compensated by calculating the average phase angle of each constellation point in the first quadrant of the constellation diagram and comparing its difference with 45 °.
Step four, down sampling
And calculating the variances of the modules of the sampling points of the signal sample at different sampling moments, sorting the variances of the modules of the sampling points at all different sampling moments, and selecting the sampling points according to the sequence from small to large of the variances by combining the down-sampling ratio.
Step five, data standardization
And calculating the mean value of the data on each dimension by using all the sample data, then subtracting the mean value from each dimension, and finally dividing each dimension of the data by the standard deviation of the data on the dimension, so that the characteristics of each dimension conform to the standard normal distribution N (0,1), and normalizing the characteristics of each dimension to the same value interval.
Step six, inputting phase data of baseband signals obtained after data preprocessing into a model network, wherein the number of sampling points after down-sampling is 1024, and designing two network modules in parallel, wherein the model structure is shown in figure 3. The specific process is as follows:
step 6.A, stacking 3 one-dimensional convolutional layers by the first module, sequentially reducing the sizes of convolutional kernels, using ReLU as an activation function of the convolutional layers, carrying out batch normalization before activation, and after the convolutional layers are stacked, following a one-dimensional global average pooling layer, flattening output to the dimension of a feature channel, reducing the number of parameters and extracting effective features;
step 6.B the second module uses LSTM layer, and divides the input signal into 4 time steps, the dimension of the hidden state is 128, a Dropout layer with 0.5 discarding probability is added, and the hidden state of the last time step is taken as the output of the module;
and step 6.C, splicing the two parallel modules in the output dimension to realize the fusion of the characteristics, and using Softmax output to perform identification and judgment after the characteristics are integrated through a full connection layer.
Step seven, signal-to-noise ratio estimation
Assuming that the noise is additive white Gaussian noise, there is
r(t)＝s(t)+n(t) (2)
Wherein
Figure BDA0002867227270000141
Figure BDA0002867227270000142
The second and fourth moments of the intermediate frequency signal can be expressed as
m2＝E[r*(t)r(t)]
m4＝E{[r*(t)r(t)]2} (4)
On the other hand, note PsAnd PnThe power of signal and noise respectively, according to the narrow-band random signal theory, for the modulation signal with constant envelope such as QPSK, there are
m2＝Ps+Pn
Figure BDA0002867227270000143
Combining the above two forms to obtain
Figure BDA0002867227270000144
The signal-to-noise ratio can be estimated according to the method.
Step eight, sample enhancement
And adding corresponding white Gaussian noise to the training data according to the estimation of the signal-to-noise ratio range of the signal sample of the test set in the step seven, so that the signal-to-noise ratio of the signal sample of the training data is consistent with that of the test set, and training the model by using the noisy training data.
The seventh step and the eighth step are preferred embodiments, and the sequence may be before training the model to optimize the training result of the model.
Nine steps, abnormity detection algorithm based on confidence score
The output of the model is a probability distribution, and the inverse of the information entropy of the output distribution is taken as a confidence score. The mathematical expression for abnormality detection by this method is as follows
Figure BDA0002867227270000151
Figure BDA0002867227270000152
Wherein x is(i)For the input of the ith test sample, hθRepresenting a trained network model, c (x)(i)) Denotes the confidence score, λ is the confidence threshold, f (x)(i)) Is an exception flag.
Step ten, selecting clustering characteristics
And outputting the 256-dimensional output of the model at the Concat layer as the clustering feature of the signal.
Step eleven, judging the optimal clustering number by an elbow method
The method comprises the steps of gradually increasing the value of the cluster number k from 1 cluster, sequentially calculating the Sum of Squares of Errors (SSE) when k takes different values, gradually reducing the value of the SSE along with the increase of k, and when the value of k is increased to the real cluster number, rapidly reducing the contribution of the increase of the value of k to the aggregation degree of each cluster and gradually smoothing in subsequent changes. The computational expression of SSE is
Figure BDA0002867227270000153
Wherein, CiDenotes the ith cluster, p denotes the cluster CiSample point of (1), miIs a cluster CiThe center of mass of the lens.
Step twelve, k-means clustering
And step 12.A, randomly initializing k points as initial cluster centers. The clustering center is a position coordinate with the same vector length as each data sample;
step 12.B, calculating the distance from each data sample to each clustering center, and for any data sample, dividing the data sample into which cluster the data sample is closest to which clustering center;
step 12.C, calculating the coordinate mean value in each cluster as the new cluster center of the cluster;
and step 12.D, repeating the steps until the center of each type does not change greatly after each iteration.
The invention is not limited to the foregoing embodiments. The invention extends to any novel feature or any novel combination of features disclosed in this specification and any novel method or process steps or any novel combination of features disclosed.

Claims (10)

1. A robustness communication radiation source intelligent identification method based on deep learning is characterized by comprising the following steps:
preprocessing signal sample data;
performing data enhancement on signal sample data by utilizing the signal-to-noise ratio of the test set sample evaluation acquired in the test environment;
performing convolution operation on signal sample data by using a CNN (convolutional neural network) module to extract signal space characteristics, extracting time domain correlation characteristics of the signal sample data by using an LSTM (linear spline) module, splicing the CNN module and the LSTM module on an output dimension, integrating the characteristics of each signal sample data through a full connection layer, and classifying the characteristics of each signal sample data by using a Softmax logistic regression model, wherein the classification process comprises the process of calculating the confidence coefficient of each signal sample data;
acquiring abnormal signal sample data according to the relation between the confidence coefficient of each signal sample data and the first condition;
and performing K-Means clustering on the characteristics of the abnormal signal sample data.
2. The robust communication radiation source intelligent identification method based on deep learning of claim 1, wherein the signal sample data preprocessing step comprises: and sequentially carrying out carrier frequency estimation, down-conversion, phase compensation, down-sampling and data standardization processing on the signal sample data.
3. The robust communication radiation source intelligent identification method based on deep learning as claimed in claim 2, wherein the data standardization process comprises: the mean of the data in each dimension is calculated using the totality of the signal sample data, subtracted in each dimension, and divided by the standard deviation of the data in that dimension.
4. The method as claimed in claim 1, wherein the CNN module stacks 3 one-dimensional convolutional layers, the sizes of convolutional kernels are sequentially reduced, the convolutional layers are activated by using a ReLU function, batch normalization is performed before activation, and after the convolutional layers are stacked, a one-dimensional global average pooling layer is followed, and an output is mapped to a feature channel dimension.
5. The robust communication radiation source intelligent identification method based on deep learning as claimed in claim 1, wherein the LSTM module divides an input signal into 4 time steps, the dimension of hidden state is 128, a Dropout layer with a discarding probability of 0.5 is added, and the hidden state of the last time step is taken as the output of the module.
6. The robust communication radiation source intelligent identification method based on deep learning as claimed in any one of claims 1 to 5, wherein the signal-to-noise ratio of the test set samples collected in the test environment is estimated by a signal-to-noise ratio estimation method based on a high-order moment.
7. The method for intelligent recognition of a robust communication radiation source based on deep learning of claim 6, wherein the data enhancement of signal sample data comprises:
white gaussian noise corresponding to the estimated signal-to-noise ratio is added to the signal sample data.
8. The method for intelligently identifying a robust communication radiation source based on deep learning according to any one of claims 1 to 5 and 7, wherein the method for calculating the confidence of each signal sample data comprises the following steps:
and taking the reciprocal of the information entropy of the probability distribution output by the Softmax logistic regression model as the confidence coefficient of the corresponding signal sample data.
9. The robust communication radiation source intelligent identification method based on deep learning of any one of claims 1 to 5 and 7, wherein the clustering cluster number of the K-Means clustering is obtained by an elbow method based on error square sum calculation.
10. The robust communication radiation source intelligent identification method based on deep learning of claim 1, 4 or 5, wherein the features of the abnormal signal sample data subjected to K-Means clustering are features extracted by a spliced CNN module and an LSTM module and integrated by a full connection layer.
CN202011592239.4A 2020-12-29 2020-12-29 Robustness communication radiation source intelligent identification method based on deep learning Pending CN112613443A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011592239.4A CN112613443A (en) 2020-12-29 2020-12-29 Robustness communication radiation source intelligent identification method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011592239.4A CN112613443A (en) 2020-12-29 2020-12-29 Robustness communication radiation source intelligent identification method based on deep learning

Publications (1)

Publication Number Publication Date
CN112613443A true CN112613443A (en) 2021-04-06

Family

ID=75248728

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011592239.4A Pending CN112613443A (en) 2020-12-29 2020-12-29 Robustness communication radiation source intelligent identification method based on deep learning

Country Status (1)

Country Link
CN (1) CN112613443A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113177520A (en) * 2021-05-26 2021-07-27 电子科技大学 Intelligent radiation source identification method based on ensemble learning
CN113255545A (en) * 2021-06-03 2021-08-13 中国人民解放军国防科技大学 Communication radiation source individual identification method combining artificial features and depth features
CN113435245A (en) * 2021-05-18 2021-09-24 西安电子科技大学 Method, system and application for identifying individual airborne radiation source

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111123039A (en) * 2019-12-31 2020-05-08 国网北京市电力公司 Power distribution network abnormal line loss diagnosis method based on synchronization characteristics and improved K-means clustering

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111123039A (en) * 2019-12-31 2020-05-08 国网北京市电力公司 Power distribution network abnormal line loss diagnosis method based on synchronization characteristics and improved K-means clustering

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
YANYAN YAO等: "Feature Extraction Method of Radiation Source in Deep Learning Based on Square Integral Bispectrum", 《JOURNAL OF PHYSICS: CONFERENCE SERIES》 *
YOUNGHWAN KIM等: "Anomaly Detection using Clustered Deep One-Class Classification", 《2020 15TH ASIA JOINT CONFERENCE ON INFORMATION SECURITY (ASIAJCIS)》 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113435245A (en) * 2021-05-18 2021-09-24 西安电子科技大学 Method, system and application for identifying individual airborne radiation source
CN113435245B (en) * 2021-05-18 2023-06-30 西安电子科技大学 Method, system and application for identifying individual aerial radiation source
CN113177520A (en) * 2021-05-26 2021-07-27 电子科技大学 Intelligent radiation source identification method based on ensemble learning
CN113177520B (en) * 2021-05-26 2022-06-28 电子科技大学 Intelligent radiation source identification method based on ensemble learning
CN113255545A (en) * 2021-06-03 2021-08-13 中国人民解放军国防科技大学 Communication radiation source individual identification method combining artificial features and depth features
CN113255545B (en) * 2021-06-03 2021-09-21 中国人民解放军国防科技大学 Communication radiation source individual identification method combining artificial features and depth features

Similar Documents

Publication Publication Date Title
CN112613443A (en) Robustness communication radiation source intelligent identification method based on deep learning
CN111562108A (en) Rolling bearing intelligent fault diagnosis method based on CNN and FCMC
CN108333468B (en) The recognition methods of bad data and device under a kind of active power distribution network
CN116866129A (en) Wireless communication signal detection method
CN111698258B (en) WiFi-based environmental intrusion detection method and system
CN102497343A (en) Combined modulation recognition method based on clustering and support vector machine
CN107392123B (en) Radio frequency fingerprint feature extraction and identification method based on coherent accumulation noise elimination
CN113542241B (en) Intrusion detection method and device based on CNN-BiGRU hybrid model
CN112347871A (en) Interference signal modulation identification method for communication carrier monitoring system
Li et al. A deep convolutional network for multitype signal detection and classification in spectrogram
Li et al. Automatic modulation classification based on bispectrum and CNN
CN117912599B (en) Food additive detection method based on artificial intelligence
CN111860447A (en) Method for accurately identifying telemetry time sequence data mode of random phase shift satellite
CN116805051A (en) Double convolution dynamic domain adaptive equipment fault diagnosis method based on attention mechanism
CN116205863A (en) Method for detecting hyperspectral image abnormal target
CN106405683B (en) Wind speed forecasting method and device based on G-L mixed noise characteristic core ridge regression technology
CN118260571A (en) Satellite signal individual identification method based on multi-mode feature fusion and deep learning
CN112751633B (en) Broadband spectrum detection method based on multi-scale window sliding
Essai et al. Automatic Modulation Classification: Convolutional Deep Learning Neural Networks Approaches
CN116346506B (en) Intelligent protection method for network security data
CN115100733B (en) RFID gesture recognition method, computer device, product and storage medium
CN107341519B (en) Support vector machine identification optimization method based on multi-resolution analysis
Wong On the use of convolutional neural networks for specific emitter identification
CN114580476A (en) Unmanned aerial vehicle signal identification model construction method and corresponding identification method and system
CN109995690A (en) The neural network self-organization method of MFSK digital signal subclass Modulation Identification

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210406