CN114826857A - OFDM signal generation method based on generation countermeasure network - Google Patents
OFDM signal generation method based on generation countermeasure network Download PDFInfo
- Publication number
- CN114826857A CN114826857A CN202210310083.9A CN202210310083A CN114826857A CN 114826857 A CN114826857 A CN 114826857A CN 202210310083 A CN202210310083 A CN 202210310083A CN 114826857 A CN114826857 A CN 114826857A
- Authority
- CN
- China
- Prior art keywords
- multiplied
- frequency
- time
- function
- discriminator
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 24
- 230000007274 generation of a signal involved in cell-cell signaling Effects 0.000 title claims abstract description 13
- 239000011159 matrix material Substances 0.000 claims abstract description 34
- 238000012549 training Methods 0.000 claims abstract description 33
- 239000013598 vector Substances 0.000 claims abstract description 16
- 238000007781 pre-processing Methods 0.000 claims abstract description 7
- 230000006870 function Effects 0.000 claims description 55
- 238000006243 chemical reaction Methods 0.000 claims description 20
- 230000004913 activation Effects 0.000 claims description 18
- 238000010606 normalization Methods 0.000 claims description 13
- 230000009977 dual effect Effects 0.000 claims description 10
- 125000004122 cyclic group Chemical group 0.000 claims description 5
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 230000009466 transformation Effects 0.000 claims description 3
- 238000004891 communication Methods 0.000 abstract description 8
- 238000005516 engineering process Methods 0.000 abstract description 3
- 238000012360 testing method Methods 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 14
- 239000000969 carrier Substances 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000004088 simulation Methods 0.000 description 2
- 101000595198 Homo sapiens Podocalyxin Proteins 0.000 description 1
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 102100036031 Podocalyxin Human genes 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000903 blocking effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000013139 quantization Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/26—Systems using multi-frequency codes
- H04L27/2601—Multicarrier modulation systems
- H04L27/2602—Signal structure
-
- 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
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Signal Processing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Digital Transmission Methods That Use Modulated Carrier Waves (AREA)
Abstract
The invention discloses an OFDM signal generation method based on a generation countermeasure network, which generates a complex protocol signal modulated by OFDM and belongs to the signal reconstruction technology in the field of communication countermeasure. Aiming at the problem that the time domain characteristics of OFDM signals are difficult to extract, firstly, FFT is used for preprocessing the time domain OFDM signals, and processed frequency domain symbol vectors are spliced into a two-dimensional data matrix; and then, storing the data matrix as a time-frequency two-dimensional pattern in a gray image mode for the training and testing of the GAN. In addition, the dual-discriminator GAN adopts a generator and two discriminators to play games simultaneously, the generator is used for generating a time-frequency two-dimensional pattern to deceive the two discriminators, and the two discriminators respectively distinguish the generated pattern and a real pattern from two aspects of subcarrier structures and constellation densities of modulation symbols so as to ensure that a generated signal meets the requirements of OFDM time-frequency domain characteristics.
Description
Technical Field
The invention belongs to a signal reconstruction technology in the field of communication countermeasure, and particularly relates to a method for generating a complex protocol signal modulated by Orthogonal Frequency Division Multiplexing (OFDM).
Background
In the face of unknown communication signals, the conventional communication countermeasure mechanism needs to perform channel estimation and signal parameter estimation, such as modulation mode, code rate, carrier frequency, frequency offset and the like, and then perform tracking type interference on the signals, or directly perform blocking type interference under the condition of failed identification of external signals. However, with the development and progress of communication and network technologies, electromagnetic signals have the characteristics of multimode, multi-system, intellectualization, high self-adaptation, rapidness, and the like, and due to the differentiation requirements of different communication scenes, a wireless communication system designs various complex protocols aiming at various scenes, and in the face of unknown complex protocol signals, the traditional interference method has the problems of difficult identification and difficult generation. In addition, when an interference signal is generated, the conventional interference method is realized by hardware modulation, different hardware systems are required to realize different complex protocol signals, and sufficient flexibility and generalization capability are lacked.
Among the modulation of various types of complex protocol signals, the OFDM modulation technique is currently most widely used, for example, the currently mainstream 4G and 5G communication protocols in mobile communication, Digital Video Broadcasting (DVB), IEEE 802.11a standard in wireless local area networks, HiperLAN standard, and the like all use OFDM modulation, but these protocols adopt different frame structures. In the face of complex protocol signals modulated by OFDM of an enemy, the traditional interference mode needs different hardware systems to reconstruct different protocols and parameter signals.
The signal generation model based on machine learning can realize the generation of signals with different parameters under a unified architecture, and has more flexibility in hardware. As an emerging Generative model, generation of a countermeasure network (GAN) can learn high-dimensional and complex real data distribution, and is widely applied to the field of data generation. The GAN usually consists of a neural network of a generator and a discriminator, and through resistance training of the generator and the discriminator, the distribution and the statistical characteristics of real sample data can be automatically learned, so that Nash equilibrium can be reached or approached under a proper condition, and at the moment, the generator of the GAN can generate data similar to the real sample. GAN finds application in data generation and signal processing in many fields, such as speech signal enhancement, music sequence generation, etc., and is very widely used in the field of computer vision, for example, image-to-image conversion, image super resolution, video generation, text generation, etc. For a set of pictures or data with better spatial correlation, it is feasible to generate data resembling real samples with GAN networks.
By utilizing the strong data generation capability of the GAN, the GAN network can be used as a uniform architecture to generate OFDM signals with different parameters. However, the OFDM symbol sequence has the characteristics of noise-like observed from the time domain, and the correlation is poor, which is not favorable for feature extraction and data distribution fitting. For a generated model, the distribution characteristics of the OFDM time domain signal are difficult to learn, and if the OFDM time domain signal is used as a training sample for training, a mapping condition that a generated signal constellation does not meet requirements is easy to occur, and the time-frequency domain parameters do not meet the requirements of a protocol frame structure, and the like, and the generated model does not have the characteristics of the OFDM signal. The characteristics of the OFDM signal need to be observed simultaneously from more dimensions, such as the time-frequency domain, so that the GAN generates a complex protocol signal that meets the required OFDM modulation.
Disclosure of Invention
In view of the existing problems, the invention provides an OFDM signal generation method based on a generation countermeasure network, by which two dimensional features of an OFDM signal in a time domain and a frequency domain can be extracted to generate an OFDM signal satisfying a specific protocol (including a subcarrier structure and a modulation mode).
In order to achieve the purpose, the invention adopts the technical scheme that:
an OFDM signal generation method based on generation countermeasure network, comprising the following steps:
a. preprocessing an input OFDM time domain signal, converting the OFDM signal into a frequency domain, normalizing the obtained frequency domain signal, splicing the time sequence of each frequency domain OFDM symbol to obtain a time-frequency two-dimensional matrix, and storing the real part and the imaginary part of the matrix in the form of a single-channel gray-scale image respectively to obtain a time-frequency two-dimensional pattern for training and verifying a generation countermeasure network;
b. b, inputting the time-frequency two-dimensional pattern obtained in the step a into a double-discriminator GAN for training, and storing a generator model after the training is converged;
c. inputting specific random noise into a trained generator model to obtain a real part or imaginary part time-frequency two-dimensional pattern of the OFDM frequency domain signal meeting the corresponding protocol;
d. and c, adding the real part and the imaginary part of the time-frequency two-dimensional pattern obtained in the step c, and performing IFFT (inverse fast Fourier transform) and parallel-serial conversion to obtain the generated OFDM time domain signal.
Further, the preprocessing in the step a comprises two steps of serial-to-parallel conversion and cyclic prefix removing fast Fourier transform.
Further, the normalization processing mode in step a is as follows:
r norm ＝r./2max(abs(r))+0.5+0.5j
wherein r represents a matrix obtained by performing serial-to-parallel conversion on the OFDM time domain signal, removing a cyclic prefix, performing FFT on each OFDM symbol, and splicing the obtained vectors according to rows; where/represents the division of each row of the matrix by the elements of the corresponding row in the column vector, max (-) represents the column vector consisting of the maximum of each row in the matrix, abs (-) represents the magnitude of each element in the matrix, and j is the imaginary unit.
Further, the dual arbiter GAN described in step b comprises a generator and two arbiters.
Further, the generator in the dual-discriminator GAN has the input of random noise and the output of a time-frequency two-dimensional pattern, and the network structure of the generator includes 9 layers, which are as follows:
the 1 st layer is an input layer, and the dimensionality of input random noise is 2 multiplied by 1;
the 2 nd to 6 th layers and the 8 th to 9 th layers are two-dimensional deconvolution layers, the zero padding is 0 x 0, the convolution kernel sizes are respectively 2 x 2, 5 x 2, 1 x 2 and 1 x 2, the step sizes are respectively 1 x 1, 2 x 2, 5 x 2, 1 x 2 and 1 x 2, except the 9 th layer, the batch normalization is carried out by using a BatchNorm function, a LeakyReLU function is used as an activation function, and the 9 th layer uses a Tanh function as an activation function;
the 7 th layer is a two-dimensional convolution layer, the zero padding is 1 multiplied by 1, the convolution kernel size is 3 multiplied by 3, the step length is 1 multiplied by 1, batch normalization is carried out by using a BatchNorm function, and a LeakyReLU function is used as an activation function;
the dimension of the output time-frequency two-dimensional pattern is 1 multiplied by 100 multiplied by 64.
Further, the input of the first discriminator in the dual discriminator GAN is a real time-frequency two-dimensional pattern and a time-frequency two-dimensional pattern generated by the generator, the output is a 1-dimensional real number, and the network structure of the first discriminator includes 7 layers, which are as follows:
the layer 1 is an input layer, and the dimension of an input time-frequency two-dimensional pattern is 1 multiplied by 100 multiplied by 64;
the 2 nd to 7 th layers are two-dimensional convolution layers, zero padding is 1 multiplied by 0,1 multiplied by 1, 1 multiplied by 0,1 multiplied by 1 and 0 multiplied by 0 respectively, convolution kernel sizes are 4 multiplied by 1, 4 multiplied by 4, 4 multiplied by 1, 4 multiplied by 4 and 2 multiplied by 2 respectively, step sizes are 2 multiplied by 1, 5 multiplied by 4, 4 multiplied by 4, 2 multiplied by 1, 2 multiplied by 2 and 1 multiplied by 1 respectively, batch normalization is carried out by using a BatchNorm function except the 7 th layer, a LeakyReLU function is used as an activation function, the 7 th layer directly outputs a number 1 or 0 as a discrimination result without using the activation function, and 1 and 0 respectively represent true and false;
the output dimension is 1 × 1 × 1.
Further, the input of the second discriminator in the dual-discriminator GAN is a constellation density map generated by the real time-frequency two-dimensional pattern through the pattern-constellation conversion network, and the input of the constellation density map generated by the time-frequency two-dimensional pattern generated by the generator through the pattern-constellation conversion network, the output is a 100-dimensional real number, and the network structure of the second discriminator includes 6 layers, specifically as follows:
the 1 st layer is an input layer, and the dimensionality of an input constellation density graph is 1 multiplied by 100;
the 2 nd to 6 th layers are two-dimensional convolution layers, zero padding is 1 multiplied by 0, and 0 multiplied by 0, convolution kernel sizes are 4 multiplied by 1, and 2 multiplied by 1, step sizes are 2 multiplied by 1, 5 multiplied by 1, 2 multiplied by 1, and 1 multiplied by 1, respectively, batch normalization is performed by using a BatchNorm function except the 6 th layer, a LeakyReLU function is used as an activation function, the 6 th layer does not use the activation function, a number 1 or 0 is directly output as a discrimination result, and 1 and 0 respectively represent true and false;
the output dimension is 1 × 1 × 100.
Further, the pattern-constellation conversion network converts the time-frequency two-dimensional pattern into a constellation density map by using a neural network, and the network structure comprises 9 layers, which are as follows:
the layer 1 is an input layer, and the dimension of an input time-frequency two-dimensional pattern is 1 multiplied by 100 multiplied by 64;
the 2 nd to 9 th layers are two-dimensional convolution layers, the output dimensions are respectively 64 multiplied by 100 multiplied by 32, 128 multiplied by 100 multiplied by 16, 256 multiplied by 100 multiplied by 8, 512 multiplied by 100 multiplied by 2, 256 multiplied by 100 multiplied by 4, 128 multiplied by 100 multiplied by 8, 64 multiplied by 100 multiplied by 20 and 1 multiplied by 100, the batch normalization is carried out by a BatchNorm function except the 9 th layer, a ReLU function is used as an activation function, and a Tanh function is used as an activation function of the 9 th layer;
the training of the pattern-constellation conversion network is carried out off-line, the optimizer uses an Adam optimizer, the learning rate is 5 multiplied by 10 -6 The input of the training is a time-frequency two-dimensional pattern, the label is a constellation density chart corresponding to the time-frequency two-dimensional pattern, and the target function during the training is MSELoss; the generation method of the constellation density chart as the training set label is as follows:
dividing [0,1] into 100 parts equally, and counting the probability density of the amplitude of each row of pixel points in the time-frequency two-dimensional pattern in the 100 intervals;
normalizing the probability density distribution obtained by each line of statistics, and respectively storing the normalized probability density distribution into vectors with dimensions of 1 multiplied by 100;
and splicing the obtained vectors according to rows to obtain a matrix with the dimensionality of 100 multiplied by 100, and storing the matrix into a single-channel gray-scale map, namely the constellation density map serving as a training label.
Further, the training in step b uses Adam optimizer, and the learning rate of the generator is 4 × 10 -4 The learning rate of the first discriminator is 1 × 10 -4 The learning rate of the second discriminator is 1 × 10 -4 The training objective function is as follows:
wherein, V (D) 1 )、V(D 2 ) V (G) is a loss function of the first discriminator, the second discriminator and the generator, G is the generator, D 1 Is a first discriminator, D 2 Is a second discriminator, C is a pattern-constellation transformation network, z is random noise input into the generator, x is a real time-frequency two-dimensional pattern as a training set, E represents expectation, and x-p data (x) Representing the probability distribution of x obeying a real dataset, z-p z (z) denotes that the noise of the input satisfies a random distribution, λ 1 And λ 2 The weights of the first and second discriminators in the generator objective function, respectively.
The invention has the beneficial effects that:
when the GAN network is trained, the dual discriminators are designed to respectively constrain a generated pattern from a time-frequency domain structure and a symbol constellation distribution, so that the generated pattern can generate OFDM signals of a subcarrier structure and a modulation mode meeting the requirements of a specific protocol.
Drawings
Fig. 1 is a schematic diagram illustrating an OFDM signal generation method based on a generation countermeasure network according to an embodiment of the present invention.
Fig. 2 is a time-frequency two-dimensional pattern corresponding to a real part of a WIFI 802.11a signal.
Fig. 3 is a network structure diagram of a dual arbiter GAN according to an embodiment of the present invention.
Fig. 4 is a schematic diagram of a conversion process of converting an OFDM real part time-frequency two-dimensional pattern with 64 subcarriers and using QPSK modulation into a constellation density chart with a resolution of 100.
Fig. 5 is a diagram of a pattern-constellation converter network structure, which takes an OFDM signal time-frequency two-dimensional pattern with 64 FFT points as an example in the embodiment of the present invention.
Fig. 6 is a schematic diagram of a generator network architecture.
Fig. 7 is a schematic diagram of a network structure of the arbiter 1 and the arbiter 2.
Fig. 8 is a time-frequency two-dimensional pattern (left diagram is real part, right side is imaginary part) generated by the generator with QPSK and 20dB snr.
Fig. 9 is a true constellation density diagram for the generated time-frequency two-dimensional pattern and the true time-frequency two-dimensional pattern (the left diagram 1 is the real part of the true pattern, the left diagram 2 is the imaginary part of the true pattern, the right diagram 2 is the real part of the generated pattern, and the right diagram 1 is the imaginary part of the generated pattern).
Fig. 10 is a comparison graph of a data segment OFDM signal time domain waveform generated based on an OFDM signal generation model for generating a countermeasure network and a data segment OFDM real signal time domain waveform of the WIFI 802.11a standard.
Detailed Description
In order to make the objects, technical means and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the following embodiments and accompanying drawings.
An OFDM signal generation method based on generation countermeasure network, the flow is as shown in figure 1, through removing CP to OFDM time domain symbol which accords with a certain specific protocol, FFT pretreatment is carried out, time-frequency two-dimensional pattern of OFDM signal is obtained, the pattern is used as real sample to be input into GAN network of multi-discriminator to train, after generator and discriminator reach Nash equilibrium, IFFT is carried out to time-frequency two-dimensional pattern generated by generator, time domain signal which also accords with the protocol can be obtained. And inputting the OFDM signals with different parameters into the system for training, and the obtained GAN network generator can realize the output of OFDM complex protocol signals with different parameters.
The method comprises the following steps:
a. as shown in fig. 1, a received OFDM time domain signal r is first preprocessed. Firstly, r is converted in series-parallel mode, CP (cyclic prefix) is removed, and the dimension N is obtained T ×N FFT R' where N T Is the number of OFDM symbols, N FFT The number of FFT points. Each row of the matrix represents a time domain sequence of one OFDM symbol. FFT is carried out on each row in the matrix r', and the signals are transformed from the time domain to the frequency domain to obtain a time-frequency two-dimensional matrix
After normalization processing shown in formula (1), a two-dimensional matrix r is obtained norm 。
Where,/represents each row of the matrix divided by the elements of the corresponding row in the column vector, max (-) represents the column vector consisting of the maximum values of each row in the matrix, abs (-) represents the magnitude of each element in the matrix, and j is an imaginary unit. Normalized matrix r norm The value range of the real part and the imaginary part of each element is [0,1]]。
Extracting the matrix r norm And each element real part and imaginary part of the matrix form a new time-frequency two-dimensional matrix respectively
And
the matrix is stored in the form of a single-channel grayscale image and used as our proposed dual-discriminator GAN training set.
Taking the OFDM signal of the data portion in the WIFI 802.11a protocol as an example, the real part time-frequency two-dimensional pattern obtained by the signal after the preprocessing is shown in fig. 2. In this protocol, the number of FFT points is 64, so the pattern has 64 columns, which respectively represent 64 frequency domain subcarriers, and 100 rows of the pattern respectively represent 100 OFDM symbols. 48 data subcarriers are arranged in 64 subcarriers, each data subcarrier carries random QPSK symbols, and black and white staggered pixel points are arranged in a real part time-frequency two-dimensional pattern; 4 sub-carriers carrying pilot frequency are white; the remaining 12 columns are empty subcarriers, which are gray in the pattern.
b. And (b) inputting the time-frequency two-dimensional pattern obtained in the step (a) into a double-discriminator GAN shown in fig. 3 for training, wherein in the double-discriminator GAN, a discriminator 2 is added in addition to the discriminator 1 in the conventional GAN. The purpose of the discriminator is to constrain the constellation distribution of the modulation symbols in the generated pattern and judge whether the modulation constellation distribution of the generated pattern is the same as the real pattern. If the two are the same, the judgment is true, and if the two are different, the judgment is false. This embodiment uses LSGAN as the base network. The labels of the false sample and the real sample are respectively represented by 0 and 1, and in order to make the generator generate the sample as real as possible, the decision boundary for the discriminator to discriminate the sample is 1. The objective function of GAN for the dual arbiter is defined as follows:
wherein G is a generator, D 1 Is a discriminator 1, D 2 For the discriminator 2, C is a pattern constellation conversion network, z is random noise of an input generator, x is a real time-frequency two-dimensional pattern used as a training set, E represents expectation, and x-p data (x) Representing x obedience real datasetProbability distribution of (1), z to p z (z) denotes that the noise of the input satisfies a random distribution, λ 1 And λ 2 The weights that the arbiter 1 and arbiter 2 occupy in the generator objective function, respectively.
In order to determine the true or false distribution of the data symbol modulation constellation in the time-frequency two-dimensional pattern, the input of the discriminator 2 is the true pattern and the constellation density map of the generated pattern respectively. The embodiment provides a pattern-constellation conversion network, which is trained offline, and has a function of realizing the conversion from a time-frequency two-dimensional pattern to a constellation density chart, and the manner of converting the time-frequency two-dimensional pattern to the constellation density chart is shown in fig. 4. Taking an OFDM signal carrying data in a WIFI 802.11a protocol as an example, the size of the input time-frequency two-dimensional pattern is 100 × 64, and represents a time-frequency two-dimensional real part pattern corresponding to a signal with a length of 100 OFDM symbol periods. Firstly, dividing [0,1] into 100 parts equally, and counting the probability density of the amplitude of each row of pixel points in the time-frequency two-dimensional pattern in the 100 intervals. And normalizing the probability density distribution obtained by statistics of each row, and respectively storing the normalized probability density distribution into vectors with the dimensionality of 1 multiplied by 100. Finally, the obtained vectors are spliced according to rows to obtain a matrix with the dimensionality of 100 multiplied by 100, and the matrix is a constellation density diagram.
This embodiment uses a similar convolutional auto-encoder to fit the quantization statistics flow described above. And extracting the constellation distribution characteristics of the time-frequency two-dimensional pattern data points through operations such as convolution, deconvolution and the like to generate a constellation density map. The structure of the pattern-constellation conversion network is shown in fig. 5.
The generator is composed of deconvolution and convolution neural network modules, and the specific structure is shown in fig. 6. Under the action of a pattern structure and constellation data dual discriminator, the input random noise is converted to approximate to a time-frequency two-dimensional pattern of a real sample. The discriminator is composed of a convolution network module, and the specific structure is shown in fig. 7. The effect is to distinguish between the true and generated samples of the input as correctly as possible.
In order to verify the effectiveness of the OFDM signal generation model based on the discriminator in this embodiment, a signal generation experiment is performed by using a data segment in a WIFI 802.11a PLCP Protocol data Unit (PCLP Protocol Date Unit, PPDU) as a real signal. The data segment is composed of OFDM symbols, the number of subcarriers is 64, wherein the number of the subcarriers carrying the data symbols is 48, and the subcarriers are respectively positioned at 2-7 th, 9-21 th, 23-27 th, 39-43 th, 45-57 th and 59-64 th subcarriers; the pilot frequency occupies 4 sub-carriers and is distributed in 8 th, 22 th, 44 th and 58 th sub-carriers; the remaining 12 are null subcarriers. The signal bandwidth is 20MHz, each OFDM symbol has a duration of 4us, the CP length is 0.8us, and in the case of a sampling rate of 20MHz, one OFDM symbol contains 80 samples, where the CP contains 16 samples. Table 1 and Table 2 show the simulation parameters and the hyper-parameters selected for model training, respectively.
TABLE 1 simulation parameters
| Parameter(s) | Value of |
| Number of data subcarriers | 48 |
| Number of pilot subcarriers | 4 |
| N FFT : total number of subcarriers/number of FFT points | 64 |
| B: bandwidth of | 20MHz |
| Modulation system | QPSK |
| SNR: signal-to-noise ratio of channel | 20dB |
| The size of the time-frequency two-dimensional graph obtained by preprocessing | 100*64 |
| Sample data set size per SNR and mode under modulation | 10000 |
TABLE 2 Dual arbiter LSGAN training hyperparameters
| Parameter(s) | Value of |
| Batch_size | 40 |
| Learning rate of generator | 4e-4 |
| Learning rate of |
1e-4 |
| Learning rate of |
1e-4 |
| | Adam |
| Arbiter | |
| 1 loss function weight λ 1 | 0.1 |
| |
0.4 |
c. Inputting random noise into the trained generator model to obtain a real part or imaginary part time-frequency two-dimensional pattern of the OFDM frequency domain signal satisfying the WIFI 802.11a PLCP protocol as shown in fig. 8, and a real constellation density diagram corresponding to the generated time-frequency two-dimensional pattern is shown in fig. 9.
d. And c, adding the real part and the imaginary part of the time-frequency two-dimensional pattern obtained in the step c, and performing IFFT and parallel-serial conversion to obtain the OFDM time domain signal meeting the WIFI 802.11a PLCP protocol as shown in the figure 10.
In summary, aiming at the problem that the time domain feature of the OFDM signal is difficult to extract, firstly, the time domain OFDM signal is preprocessed by using Fast Fourier Transform (FFT), and the processed frequency domain symbol vectors are spliced into a two-dimensional data matrix; and then, storing the data matrix as a time-frequency two-dimensional pattern in a gray image mode for the training and testing of the GAN. In addition, the invention designs double discriminators GAN, uses a generator and two discriminators to play games simultaneously, the generator aims to generate a time-frequency two-dimensional pattern to deceive the two discriminators, and the two discriminators respectively distinguish the generated pattern and a real pattern from the two aspects of subcarrier structures and constellation densities of modulation symbols so as to ensure that a generated signal meets the requirements of OFDM time-frequency domain characteristics. The invention can realize the generation of complex protocol OFDM signals.
Claims (9)
1. An OFDM signal generation method based on a generation countermeasure network, characterized by comprising the following steps:
a. preprocessing an input OFDM time domain signal, converting the OFDM signal into a frequency domain, normalizing the obtained frequency domain signal, splicing the time sequence of each frequency domain OFDM symbol to obtain a time-frequency two-dimensional matrix, and storing the real part and the imaginary part of the matrix in the form of a single-channel gray-scale image respectively to obtain a time-frequency two-dimensional pattern for training and verifying a generation countermeasure network;
b. b, inputting the time-frequency two-dimensional pattern obtained in the step a into a double-discriminator GAN for training, and storing a generator model after the training is converged;
c. inputting specific random noise into a trained generator model to obtain a real part or imaginary part time-frequency two-dimensional pattern of the OFDM frequency domain signal meeting the corresponding protocol;
d. and c, adding the real part and the imaginary part of the time-frequency two-dimensional pattern obtained in the step c, and performing IFFT (inverse fast Fourier transform) and parallel-serial conversion to obtain the generated OFDM time domain signal.
2. The OFDM signal generation method as claimed in claim 1, wherein the preprocessing in step a comprises serial-to-parallel conversion and cyclic prefix removal fast fourier transform.
3. The method as claimed in claim 1, wherein the normalization in step a is performed by:
r norm ＝r./2max(abs(r))+0.5+0.5j
wherein r represents a matrix obtained by performing serial-to-parallel conversion on the OFDM time domain signal, removing a cyclic prefix, performing FFT on each OFDM symbol, and splicing the obtained vectors according to rows; where/represents the division of each row of the matrix by the elements of the corresponding row in the column vector, max (-) represents the column vector consisting of the maximum of each row in the matrix, abs (-) represents the magnitude of each element in the matrix, and j is the imaginary unit.
4. The OFDM signal generating method as claimed in claim 1, wherein the dual arbiter GAN in step b comprises a generator and two arbiters.
5. The method as claimed in claim 4, wherein the generator in the dual-arbiter GAN has an input of random noise and an output of time-frequency two-dimensional pattern, and the network structure of the generator comprises 9 layers, specifically as follows:
the 1 st layer is an input layer, and the dimensionality of input random noise is 2 multiplied by 1;
the 2 nd to 6 th layers and the 8 th to 9 th layers are two-dimensional deconvolution layers, the zero padding is 0 x 0, the convolution kernel sizes are respectively 2 x 2, 5 x 2, 1 x 2 and 1 x 2, the step sizes are respectively 1 x 1, 2 x 2, 5 x 2, 1 x 2 and 1 x 2, except the 9 th layer, the batch normalization is carried out by using a BatchNorm function, a LeakyReLU function is used as an activation function, and the 9 th layer uses a Tanh function as an activation function;
the 7 th layer is a two-dimensional convolution layer, the zero padding is 1 multiplied by 1, the convolution kernel size is 3 multiplied by 3, the step length is 1 multiplied by 1, batch normalization is carried out by using a BatchNorm function, and a LeakyReLU function is used as an activation function;
the dimension of the output time-frequency two-dimensional pattern is 1 multiplied by 100 multiplied by 64.
6. The method as claimed in claim 4, wherein the input of the first discriminator in the dual discriminator GAN is a real time-frequency two-dimensional pattern and the time-frequency two-dimensional pattern generated by the generator, and the output is a 1-dimensional real number, and the network structure of the first discriminator includes 7 layers, which are as follows:
the layer 1 is an input layer, and the dimension of an input time-frequency two-dimensional pattern is 1 multiplied by 100 multiplied by 64;
the 2 nd to 7 th layers are two-dimensional convolution layers, zero padding is 1 multiplied by 0,1 multiplied by 1, 1 multiplied by 0,1 multiplied by 1 and 0 multiplied by 0 respectively, convolution kernel sizes are 4 multiplied by 1, 4 multiplied by 4, 4 multiplied by 1, 4 multiplied by 4 and 2 multiplied by 2 respectively, step sizes are 2 multiplied by 1, 5 multiplied by 4, 4 multiplied by 4, 2 multiplied by 1, 2 multiplied by 2 and 1 multiplied by 1 respectively, batch normalization is carried out by using a BatchNorm function except the 7 th layer, a LeakyReLU function is used as an activation function, the 7 th layer directly outputs a number 1 or 0 as a discrimination result without using the activation function, and 1 and 0 respectively represent true and false;
the output dimension is 1 × 1 × 1.
7. The method according to claim 4, wherein the inputs of the second discriminator in the dual discriminator GAN are a constellation density map generated by the real time-frequency two-dimensional pattern through the pattern-constellation conversion network and a constellation density map generated by the time-frequency two-dimensional pattern generated by the generator through the pattern-constellation conversion network, and the output is a real 100-dimensional number, and the network structure of the second discriminator includes 6 layers, which are as follows:
the 1 st layer is an input layer, and the dimensionality of an input constellation density graph is 1 multiplied by 100;
the 2 nd to 6 th layers are two-dimensional convolution layers, zero padding is 1 multiplied by 0, and 0 multiplied by 0, convolution kernel sizes are 4 multiplied by 1, and 2 multiplied by 1, step sizes are 2 multiplied by 1, 5 multiplied by 1, 2 multiplied by 1, and 1 multiplied by 1, respectively, batch normalization is performed by using a BatchNorm function except the 6 th layer, a LeakyReLU function is used as an activation function, the 6 th layer does not use the activation function, a number 1 or 0 is directly output as a discrimination result, and 1 and 0 respectively represent true and false;
the output dimension is 1 × 1 × 100.
8. The method according to claim 7, wherein the pattern-constellation transformation network transforms the time-frequency two-dimensional pattern into the constellation density map by using a neural network, and the network structure comprises 9 layers, which are as follows:
the layer 1 is an input layer, and the dimension of an input time-frequency two-dimensional pattern is 1 multiplied by 100 multiplied by 64;
the 2 nd to 9 th layers are two-dimensional convolution layers, the output dimensions are respectively 64 multiplied by 100 multiplied by 32, 128 multiplied by 100 multiplied by 16, 256 multiplied by 100 multiplied by 8, 512 multiplied by 100 multiplied by 2, 256 multiplied by 100 multiplied by 4, 128 multiplied by 100 multiplied by 8, 64 multiplied by 100 multiplied by 20 and 1 multiplied by 100, the batch normalization is carried out by a BatchNorm function except the 9 th layer, a ReLU function is used as an activation function, and a Tanh function is used as an activation function of the 9 th layer;
the training of the pattern-constellation conversion network is carried out off-line, the optimizer uses an Adam optimizer, the learning rate is 5 multiplied by 10 -6 The input of the training is a time-frequency two-dimensional pattern, the label is a constellation density chart corresponding to the time-frequency two-dimensional pattern, and the target function during the training is MSELoss; the generation method of the constellation density chart as the training set label is as follows:
dividing [0,1] into 100 parts equally, and counting the probability density of the amplitude of each row of pixel points in the time-frequency two-dimensional pattern in the 100 intervals;
normalizing the probability density distribution obtained by each line statistics, and respectively storing the normalized probability density distribution into vectors with dimensions of 1 multiplied by 100;
and splicing the obtained vectors according to rows to obtain a matrix with the dimensionality of 100 multiplied by 100, and storing the matrix into a single-channel gray-scale map, namely the constellation density map serving as a training label.
9. The method as claimed in claim 8, wherein the training in step b uses Adam optimizer, and the learning rate of the generator is 4 x 10 -4 The learning rate of the first discriminator is 1 × 10 -4 The learning rate of the second discriminator is 1 × 10 -4 The training objective function is as follows:
wherein, V (D) 1 )、V(D 2 ) V (G) is a loss function of the first discriminator, the second discriminator and the generator, G is the generator, D 1 Is a first discriminator, D 2 Is a second discriminator, C is a pattern-constellation transformation network, z is random noise input into the generator, x is a real time-frequency two-dimensional pattern as a training set, E represents expectation, and x-p data (x) Representing the probability distribution of x obeying a real dataset, z-p z (z) denotes that the noise of the input satisfies a random distribution, λ 1 And λ 2 In generator objective function for first and second discriminators respectivelyThe occupied weight.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210310083.9A CN114826857B (en) | 2022-03-28 | 2022-03-28 | OFDM signal generation method based on generation countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210310083.9A CN114826857B (en) | 2022-03-28 | 2022-03-28 | OFDM signal generation method based on generation countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114826857A true CN114826857A (en) | 2022-07-29 |
| CN114826857B CN114826857B (en) | 2024-05-03 |
Family
ID=82531444
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210310083.9A Active CN114826857B (en) | 2022-03-28 | 2022-03-28 | OFDM signal generation method based on generation countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114826857B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115905819A (en) * | 2023-03-09 | 2023-04-04 | 中国民用航空飞行学院 | rPPG signal generation method and device based on generation countermeasure network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111178260A (en) * | 2019-12-30 | 2020-05-19 | 山东大学 | Modulation signal time-frequency diagram classification system based on generation countermeasure network and operation method thereof |
| US20200408864A1 (en) * | 2019-06-26 | 2020-12-31 | Siemens Healthcare Gmbh | Progressive generative adversarial network in medical image reconstruction |
| CN112804048A (en) * | 2021-04-12 | 2021-05-14 | 南京信息工程大学 | Physical layer chaotic encryption optical transmission method based on generation countermeasure network |
| CN114065841A (en) * | 2021-10-26 | 2022-02-18 | 中国电子科技集团公司第五十四研究所 | Channel characteristic migration method based on star-type generation countermeasure network |
-
2022
- 2022-03-28 CN CN202210310083.9A patent/CN114826857B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200408864A1 (en) * | 2019-06-26 | 2020-12-31 | Siemens Healthcare Gmbh | Progressive generative adversarial network in medical image reconstruction |
| CN111178260A (en) * | 2019-12-30 | 2020-05-19 | 山东大学 | Modulation signal time-frequency diagram classification system based on generation countermeasure network and operation method thereof |
| CN112804048A (en) * | 2021-04-12 | 2021-05-14 | 南京信息工程大学 | Physical layer chaotic encryption optical transmission method based on generation countermeasure network |
| CN114065841A (en) * | 2021-10-26 | 2022-02-18 | 中国电子科技集团公司第五十四研究所 | Channel characteristic migration method based on star-type generation countermeasure network |
Non-Patent Citations (1)
| Title |
|---|
| 杨鸿杰, 陈丽, 张君毅: "基于生成对抗网络的数字信号生成技术研究", 电子测量技术, vol. 43, no. 20, 31 October 2020 (2020-10-31) * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115905819A (en) * | 2023-03-09 | 2023-04-04 | 中国民用航空飞行学院 | rPPG signal generation method and device based on generation countermeasure network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114826857B (en) | 2024-05-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Shi et al. | Deep learning-based automatic modulation recognition method in the presence of phase offset | |
| Dileep et al. | Dense layer dropout based CNN architecture for automatic modulation classification | |
| Yang et al. | Deep learning aided method for automatic modulation recognition | |
| Dai et al. | Automatic modulation classification using stacked sparse auto-encoders | |
| CN110175560A (en) | A kind of radar signal intra-pulse modulation recognition methods | |
| CN109617843B (en) | KNN-based elastic optical network modulation format identification method | |
| Wang et al. | Signal detection in uplink time-varying OFDM systems using RNN with bidirectional LSTM | |
| Park et al. | Deep learning-based automatic modulation classification with blind OFDM parameter estimation | |
| CN114826857B (en) | OFDM signal generation method based on generation countermeasure network | |
| CN113726711B (en) | OFDM receiving method and device, and channel estimation model training method and device | |
| CN112242969A (en) | Novel single-bit OFDM receiver based on model-driven deep learning | |
| Cheng et al. | A ResNet-DNN based channel estimation and equalization scheme in FBMC/OQAM systems | |
| Zhang et al. | Label-assisted transmission for short packet communications: A machine learning approach | |
| Kumar et al. | Automatic modulation classification for adaptive OFDM systems using convolutional neural networks with residual learning | |
| Hao et al. | Frequency domain analysis and convolutional neural network based modulation signal classification method in OFDM system | |
| CN117633656A (en) | Radar signal identification method based on time-frequency analysis and improved convolutional neural network | |
| He et al. | Deep learning-based automatic modulation recognition algorithm in non-cooperative communication systems | |
| An et al. | Blind high-order modulation recognition for beyond 5G OSTBC-OFDM systems via projected constellation vector learning network | |
| CN117640300A (en) | Channel estimation and demodulation method based on machine learning | |
| CN115913850B (en) | Open set modulation identification method based on residual error network | |
| An et al. | Blind multicarrier waveform recognition based on spatial-temporal learning neural networks | |
| Zhang et al. | Deep learning‐based digital signal modulation identification under different multipath channels | |
| Yıldırım et al. | Deep receiver design for multi-carrier waveforms using cnns | |
| Jingpeng et al. | Modulation recognition algorithm using innovative CNN and cyclic-spectrum graph | |
| CN114298113A (en) | Internet of things-oriented dual-path machine learning modulation mode identification method |
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 |