CN113325375B - An Adaptive Cancellation Method Based on Deep Neural Network - Google Patents

Abstract

The present invention provides an adaptive cancellation method based on a deep neural network, 1) defining a signal model received by a receiving antenna, including transmitting signal power P _f , power amplifier nonlinear distortion function G[ ] and carrier center frequency f _c ; 2) Define the model of the nonlinear power amplifier; 3) Perform nonlinear modeling on the target signal, and use a large amount of data to train the DNN network; 4) Input the signal generated by the original reference signal through the trained network as a new reference signal Adaptive filter; 5) Comparing the signals before and after the adaptive filter cancels. The invention uses a large amount of training prior information to simulate the non-linear characteristics of the power amplifier of the radar jammer to solve the interference problem. This method directly estimates the amplitude of the signal and uses a large amount of data to reduce the algorithm steps.

Description

An Adaptive Cancellation Method Based on Deep Neural Network

technical field

The invention belongs to the field of self-interference cancellation of radar jammers, in particular to an adaptive cancellation method based on a deep neural network.

Background technique

The elimination of self-coupling interference in radar jammers has always been a key technology and hot topic. In recent years, adaptive algorithms have been widely used to eliminate self-coupling interference. With the increasing complexity of the electromagnetic environment, self-interference signals are also increasingly difficult to estimate. It is difficult for traditional algorithms to adapt to the nonlinear signals generated by the power amplifier of radar jammers and effectively eliminate self-interference signals.

With the continuous deepening of people's research on adaptive algorithms, the application of adaptive algorithms in radar jammers is becoming more and more mature. At present, some adaptive filtering algorithms are mainly applied to eliminate self-interference signals, such as the normalized least mean square error algorithm (Normalized Least Mean Square Error, NLMS for short). D.Lee and B.Min("Results and trade-off of self-interference cancellation in a full-duplexradio front-end,"2015 International Workshop on Antenna Technology(iWAT),Seoul,pp.249-251,2015.) It is proved that with the adaptive change of the weight of the adaptive filter, the error signal will be closer to the target value. This method only has good experimental results for radio frequency cancellation in indoor electromagnetic environments, and does not consider the nonlinearity of self-interference signals. L.Sun, Y.Li, Y.Zhao, L.Huang and Z.Gao("Optimized adaptive algorithm of digital self-interference cancellation based on improved variable step," 2015 IEEE 9th International Conference on Anti-counterfeiting, Security, and Identification( ASID), Xiamen, pp.176-179, 2015.) proposed an adaptive digital self-interference cancellation optimization algorithm based on improved variable step size, using the iterative threshold to establish a relationship between the step factor and the error signal A new nonlinear relationship overcomes the problem that the error signal changes slowly when the error signal approaches zero, and accelerates the convergence speed. This method does not mention the cancellation of the nonlinear self-jamming signal of the radar jammer. Dani Korpi, Lauri Anttila, and Mikko Valkama ("Nonlinear self-interference cancellation in MIMO full-duplex transceivers under crosstalk." Eurasip Journal on Wireless Communications & Networking 2017.1 (2017).) proposed a method for in-band multiple-input multiple-output (MIMO) A new digital self-interference canceller for full-duplex radios. It introduces various full-duplex models in detail, including the analysis of nonlinear components of self-interference signals, but does not propose how to eliminate nonlinear self-interference signals. To sum up, the above cancellation methods are aimed at the relatively simple electromagnetic environment, and the generated self-interference signal has a better cancellation effect when it is linear, and when the nonlinear self-interference generated by the power amplifier of the radar jammer When using these methods to achieve cancellation results, it is not easy to obtain good cancellation results.

Since the neural network can solve nonlinear problems very well, and recent studies by some scholars have shown that the neural network can be used for channel modeling in full-duplex communication, so that the reconstructed channel signal is more accurate. This method is based on adaptive filtering, and the weight of the filter is estimated by DNN (Deep Neural Network), but the nonlinear performance of the adaptive filter itself is not particularly good, so DNN is introduced to optimize the performance of the adaptive filter.

Contents of the invention

The invention proposes an adaptive cancellation method based on a deep neural network, which is used to solve the problem that the traditional adaptive cancellation algorithm has poor cancellation effect when dealing with self-interference signals formed by nonlinear power amplifiers. The invention uses a large amount of training prior information to simulate the nonlinear characteristics of the power amplifier of the radar jammer to solve the interference problem. This method directly estimates the amplitude of the signal and uses a large amount of data to reduce the algorithm steps.

The purpose of the present invention is achieved like this:

A kind of self-adaptive cancellation method based on deep neural network, comprises the steps:

Step 1: Define the signal model received by the receiving antenna, including the transmitted signal power P _f , the nonlinear distortion function G[ ] of the power amplifier and the carrier center frequency f _c ;

The signal received by the receiving antenna includes target signal, noise signal and self-interference signal; under normal circumstances, the noise signal is usually zero-mean Gaussian white noise, represented by n(t), and its power is P _n (ω), which is

Define the expected target signal r(t) received by the receiving antenna as:

Among them, P _f is the power of the radar transmitted signal, G[ ] represents the nonlinear distortion function of the power amplifier, d ^f (t) represents the modulated baseband waveform, and f _c represents the center frequency of the carrier;

The self-interference signal in the radar jammer is the delay of the transmitted signal, so the self-interference signal x(t) can be set as the following formula:

x(t)＝r(t-τ ₁ )+r(t-τ ₂ )+·····+r(t-τ _n ) (3)

Among them, τ ₁ , τ ₂ and τ _n are the interference delays in simulated actual situations;

After the target signal is modulated by the baseband, d ⁿ (t) is obtained by x(t), and x _PA (t) is obtained by power amplification; finally, it is sent by the transmitting antenna, and the transmitting signal will inevitably be input by the receiving antenna to form a self-interference signal SI( t):

P _n The power of the signal transmitted by the jammer, f _c represents the center frequency of the carrier, φ represents the initial phase of the carrier, and k represents the known coefficient;

Therefore, the signal y(t) model actually received by the receiving antenna is:

y(t)=r(t)+SI(t)+n(t) (5)

Step 2: Define the model of the nonlinear power amplifier;

The main cause of the nonlinear distortion of the power amplifier is AM/AM distortion. AM/AM distortion refers to the distortion of the amplitude of the output signal caused by the amplitude change of the input signal; the power amplifier used is a traveling wave tube amplifier, and its nonlinear distortion can be used Described by the Saleh model, the characteristic functions of AM/AM and AM/PM of the Saleh model are:

Among them, r is the magnitude of the input signal; α _a , β _a , α _φ and β _φ are the model parameters; a suitable fixed model is obtained by adjusting the four parameters; the nonlinear characteristics of the Saleh model include the signal passing through the Saleh model The subsequent amplitude and phase changes;

Taking the derivative of formula (6), it is obtained that when the input signal reaches the input saturation amplitude of the amplifier model, the input signal amplitude is:

The model obtains the maximum output signal amplitude:

f(A) _max = α _a A _sat /2 (9)

The maximum output value of the power amplifier determines the maximum value that can be corrected by linearization; if the linear output value corresponding to the input amplitude value of the power amplifier is greater than the maximum output amplitude value of the power amplifier, the nonlinear distortion cannot be compensated;

Step 3: Perform nonlinear modeling of the target signal, and use a large amount of data to train the DNN network;

3a) The training data are in the form of two kinds of signals: Linear Frequency Modulation (LFM) signal and BPSK signal as the target signal; for these two signals, two data sets are made respectively, each data set contains 10000 samples; each The LFM sample is a pulse signal with a pulse width of 3 μs, each BPSK sample contains 13 symbols, and the sampling points of each symbol at the sampling frequency are 70 points; for the LFM signal, define the LFM signal sampling frequency f _s =300MHz , the carrier frequency is 100MHz, the bandwidth is 20MHz, and the amplitude of the signal is 20; the sampling frequency of the definition BPSK signal is f _s =300MHz, the amplitude is 20, and the carrier frequency is 50MHz;

3b) The target signal generated in 3a is the label of the training sample obtained through the Saleh model in step 2 (the saturation output amplitude of the model is 80); wherein, the parameters of the Saleh model are set as follows α _a =2, α _φ = π/3, β _a = 1.5625e ^-5 , β _φ = 1; the generated target signal is used as the training signal of the DNN network, and the signal passed through the Saleh model is used as the label of the training data; after nonlinear amplification, the data in Significant changes have taken place in both the time and frequency domains;

3c) The input layer of the DNN network is x(t), the number of nodes in the hidden layer is enlarged and reduced respectively, the final number of nodes in the output layer is the same as the input layer, and the output signal x _AP (t) after nonlinear amplification is obtained; in the experiment Among them, the hidden layer uses the Relu (rectified linear units) activation function, and the hidden layers of the deep neural network are Num, 1024, 2048, ..., 1024, Num; where Num is the number of points for each sample, and the specific number of layers can be adjusted; The loss quadratic cost function in a neural network is:

Loss＝mean(square(xx _AP )) (10)

The output of each node of the DNN network is a nonlinear activation function applied to its input; the weights between layers in the neural network are optimized through a large amount of learning, and the expected output of training samples containing known inputs is learned;

Step 4: Input the signal generated by the original reference signal through the trained network as a new reference signal into the adaptive filter;

The signal received by the receiver includes target signal, noise signal and self-interference signal. The sum of these signals is the signal before we cancel. The self-interference signal SI(t) in formula (5) is the target to be eliminated;

4a) Taking the LFM signal as an example, the received signal becomes y(n) after denoising and sampling, and y(n) is the cancellation signal input to the adaptive filter, and the duration is 3 μs.

4b) Delay the y(n) signal in the first step by 0.5 μs, and then use the Saleh model in step 2 to perform nonlinear processing to simulate the nonlinear characteristics of the power amplifier, and use the processed self-interference signal SI(n ), adding SI(n) to y(n) and Gaussian white noise as the target cancellation signal;

4c) Input the y(n) signal delayed by 0.5 μs in the second step into the DNN neural network as a training sample, and use the signal after nonlinear processing of the Saleh model as a label, and the signal x _AP estimated by the deep neural network (n) as a reference signal for an adaptive filter;

4d) The signals generated in steps 4b and 4c are input to the adaptive filter of the LMS algorithm, and the output signal x(n) after the adaptive filter cancellation is obtained, and the result of LFM signal cancellation is achieved;

Changing the LFM signal in the above steps into a BPSK signal can also achieve the cancellation result.

Compared with prior art, the beneficial effect of the present invention is:

The invention proposes a jammer signal interference elimination method based on a deep neural network. On the basis of the existing interference elimination theory, the nonlinear characteristics of the power amplifier of the radar jammer are considered, and the nonlinearity is introduced into the modeling of the interference signal. Through experimental simulation, if the training data is rich enough and the sampling frequency is not high, the nonlinear components in the self-jamming signal received by the jammer generated by the nonlinear characteristics of the power amplifier can be eliminated.

Description of drawings

Figure 1 is a DNN-based self-interference cancellation system model.

Figure 2 shows the nonlinear characteristics of the Saleh model.

Figure 3 is the LFM signal before and after nonlinear amplification

Figure 4 is the BPSK signal before and after nonlinear amplification.

Figure 5 is a structural diagram of the DNN network model.

Figure 6 is the result of LFM signal cancellation based on the DNN method.

Figure 7 is the result of BPSK signal cancellation based on the DNN method.

Fig. 8 is the LFM signal cancellation result of the traditional method.

detailed description

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

The technical solution of this invention is a DNN-based adaptive cancellation algorithm, which includes the following steps:

1) Define the signal model received by the receiving antenna, including the transmitted signal power P _f , the nonlinear distortion function G[·] of the power amplifier and the carrier center frequency f _c .

2) Define the model of the nonlinear power amplifier.

3) The target signal is nonlinearly modeled, and a large amount of data is used to train the DNN network.

4) The signal generated after the original reference signal passes through the trained network is input into the adaptive filter as a new reference signal.

5) Compare the signals before and after the adaptive filter cancellation.

Step 1: Define the signal model received by the receiving antenna, including the transmitted signal power P _f , the nonlinear distortion function G[·] of the power amplifier and the carrier center frequency f _c .

The received signal of the receiving antenna includes target signal, noise signal and self-interference signal. In normal situations, the noise signal is usually Gaussian white noise with zero mean value, denoted by n(t), and its power is P _n (ω), namely

Define the expected target signal r(t) received by the receiving antenna as:

Among them, P _f is the power of the radar transmitted signal, G[·] represents the nonlinear distortion function of the power amplifier, d ^f (t) represents the modulated baseband waveform, and f _c represents the center frequency of the carrier.

The self-interference signal in the radar jammer is the delay of the transmitted signal, so the self-interference signal x(t) can be set as the following formula:

x(t)＝r(t-τ ₁ )+r(t-τ ₂ )+·····+r(t-τ _n ) (3)

Among them, τ ₁ , τ ₂ and τ _n are interference delays simulated in actual situations.

After the baseband modulation of the target signal, d ⁿ (t) is obtained by x(t), and x _PA (t) is obtained by power amplification. Finally, it is sent by the transmitting antenna, and the transmitting signal will inevitably be input into the self-interference signal SI(t) formed by the receiving antenna:

P _n is the power of the signal transmitted by the jammer, f _c represents the center frequency of the carrier, φ represents the initial phase of the carrier, and k represents the known coefficient.

Therefore, the signal y(t) model actually received by the receiving antenna is:

y(t)=r(t)+SI(t)+n(t) (5)

Step 2: Define the model of the nonlinear power amplifier.

The cause of the nonlinear distortion of the power amplifier is mainly AM/AM distortion, which refers to the distortion of the amplitude of the output signal caused by the amplitude change of the input signal. The power amplifier that the present invention adopts is a traveling wave tube amplifier, and its nonlinear distortion can be described by Saleh model, and the characteristic function of AM/AM and AM/PM of Saleh model is:

where r is the magnitude of the input signal. α _a , β _a , α _φ and β _φ are model parameters. A suitable fixed model can be obtained by adjusting four parameters. Figure 2 shows the nonlinear characteristics of the Saleh model, which includes the amplitude and phase changes of the signal after passing through the Saleh model.

Taking the derivative of formula (6), it can be obtained that when the input signal reaches the input saturation amplitude of the amplifier model, the input signal amplitude is:

The model obtains the maximum output signal amplitude:

f(A) _max = α _a A _sat /2 (9)

The maximum output value of the power amplifier determines the maximum value that the linearization can correct. If the linear output value corresponding to the input amplitude value of the power amplifier is greater than the maximum output amplitude value of the power amplifier, the nonlinear distortion cannot be compensated.

Step 3: The target signal is nonlinearly modeled, and a large amount of data is used to train the DNN network.

3a) The training data are respectively in the form of two signals: linear frequency modulation (LFM) signal and BPSK signal as the target signal. For these two signals, two datasets were made respectively, each containing 10000 samples. Each LFM sample is a pulse signal with a pulse width of 3 μs, each BPSK sample contains 13 symbols, and the sampling points of each symbol at the sampling frequency are 70 points. For LFM signal, define LFM signal sampling frequency f _s =300MHz, carrier frequency 100MHz, bandwidth 20MHz, signal amplitude 20; define BPSK signal sampling frequency f _s =300MHz, amplitude 20, carrier frequency 50MHz.

3b) The target signal generated in 3a is the label of the training samples obtained by the Saleh model in step 2 (the saturation output amplitude of the model is 80). Wherein, the parameters of the Saleh model are set as follows α _a =2, α _φ =π/3, β _a =1.5625e ⁻⁵ , β _φ =1. The generated target signal is used as the training signal of the DNN network, and the signal passed through the Saleh model is used as the label of the training data. As shown in Figure 3 and Figure 4, after nonlinear amplification, the data has obvious changes in both the time domain and the frequency domain.

3c) Figure 5 is a structural diagram of the DNN network model. The input layer of the DNN network is x(t), the number of nodes in the hidden layer is enlarged and reduced respectively, and the final number of nodes in the output layer is the same as that in the input layer, and the non-linearly enlarged The output signal x _AP (t). In the experiment, the hidden layer uses the Relu (rectified linear units) activation function, and the hidden layers of the deep neural network are Num, 1024, 2048, ..., 1024, Num. Among them, Num is the number of points for each sample, and the specific number of layers can be adjusted. The loss quadratic cost function in a neural network is:

Loss＝mean(square(xx _AP )) (10)

The output of each node in a DNN network is a nonlinear activation function applied to its input. The weights between layers in a neural network are optimized through extensive learning and learn the expected output from training samples containing known inputs.

Step 4: Input the signal generated by the original reference signal through the trained network as a new reference signal into the adaptive filter.

Figure 1 is a block diagram of the adaptive filter system. The signal received by the receiver includes the target signal, noise signal and self-interference signal. The sum of these signals is the signal before we cancel, such as formula (5), where the self-interference signal SI(t) is the target to eliminate.

4a) Taking the LFM signal as an example, the received signal becomes y(n) after denoising and sampling. As shown in the figure, y(n) is the cancellation signal input to the adaptive filter, and the duration is 3 μs.

4b) Delay the y(n) signal in the first step by 0.5 μs, and then use the Saleh model in step 2 to perform nonlinear processing to simulate the nonlinear characteristics of the power amplifier, and use the processed self-interference signal SI(n ), adding SI(n) to y(n) and Gaussian white noise as the target cancellation signal.

4c) Input the y(n) signal delayed by 0.5 μs in the second step into the DNN neural network as a training sample, and use the signal after nonlinear processing of the Saleh model as a label, and the signal x _AP estimated by the deep neural network (n) As a reference signal for the adaptive filter.

4d) The signals generated in steps 4b and 4c are input to the adaptive filter of the LMS algorithm to obtain the output signal x(n) after the adaptive filter cancels, and the result of the LFM signal cancellation is shown in Figure 6.

Similarly, change the LFM signal in the above steps into a BPSK signal, and the cancellation result is shown in Figure 7.

Step 5: Contrast with traditional adaptive methods.

Figure 6 shows the cancellation result of the new method for the LFM signal, and Figure 8 shows the cancellation result of the existing LMS algorithm. Through comparison, it can be clearly seen that the DNN-based method has a better cancellation effect on the self-interference signal. Compared with the traditional adaptive algorithm, it is found that the new method is used to eliminate the self-interference signal, and the BPSK signal and the LFM signal have similar cancellation results.

Claims (1)

1. A self-adaptive cancellation method based on a deep neural network is characterized by comprising the following steps:

step 1: defining a model of the signal received by the receiving antenna, including the power P of the transmitted signal _f Nonlinear distortion function G [ DEG ] of power amplifier]And carrier center frequency f _c ；

The received signals of the receiving antenna comprise target signals, noise signals and self-interference signals; in the normal case, the noise signal is usually white Gaussian noise with zero mean, denoted by n (t), and has a power P _n (ω), i.e.

Defining the expected target signal r (t) received by the receiving antenna as:

wherein, P _f Is the power of the radar transmitted signal, G [. Cndot.)]Representing the nonlinear distortion function of the power amplifier, d ^f (t) denotes a modulated baseband waveform, f _c Represents a center frequency of the carrier;

the self-interference signal in the radar jammer is a delay of the transmitted signal, so the self-interference signal x (t) can be set as follows:

x(t)＝r(t-τ ₁ )+r(t-τ ₂ )+······+r(t-τ _n ) (3)

wherein, tau ₁ 、τ ₂ And τ _n Is to simulate the interference delay in the actual situation;

after the target signal is modulated by baseband, d ⁿ (t) is obtained from x (t), x _PA (t) is obtained by power amplification; finally by transmitting antennasSending out, the transmission signal will inevitably input the self-interference signal SI (t) formed by the receiving antenna:

P _n power of the signal transmitted by the jammer, f _c Represents the center frequency of the carrier, phi represents the initial phase of the carrier, and k represents a known coefficient;

therefore, the signal y (t) actually received by the receiving antenna is modeled as:

y(t)＝r(t)+SI(t)+n(t) (5)

step 2: defining a model of the non-linear power amplifier;

the reason for causing the nonlinear distortion of the power amplifier is mainly AM/AM distortion, wherein the AM/AM distortion refers to the distortion of the amplitude of an output signal caused by the amplitude change of an input signal; the adopted power amplifier is a traveling wave tube amplifier, the nonlinear distortion of the traveling wave tube amplifier can be described by a Talbot model, and the characteristic functions of AM/AM and AM/PM of the Talbot model are as follows:

where r is the amplitude of the input signal; alpha is alpha _a 、β _a 、α _φ And beta _φ Is a model parameter; obtaining a proper fixed model by adjusting the four parameters; the nonlinear characteristics of the salech model comprise amplitude and phase changes of a signal after the signal passes through the salech model;

taking the derivative of equation (6) to obtain the input signal amplitude when the input signal reaches the input saturation amplitude of the amplifier model:

the model obtains the maximum output signal amplitude:

f(A) _max ＝α _a A _sat /2 (9)

the maximum output value of the power amplifier determines the maximum value that the linearization can correct; if the linear output value corresponding to the input amplitude value of the power amplifier is larger than the maximum output amplitude value of the power amplifier, the nonlinear distortion cannot be compensated;

and 3, step 3: carrying out nonlinear modeling on a target signal, and training a DNN network by using a large amount of data;

3a) The training data takes the form of two signals, namely a Linear Frequency Modulation (LFM) signal and a BPSK signal, as target signals respectively; for these two signals, two data sets were made, each containing 10000 samples; each LFM sample is a pulse signal with the pulse width of 3 mu s, each BPSK sample comprises 13 symbols, and the number of sampling points of each symbol at the sampling frequency is 70; for the LFM signal, the LFM signal sampling frequency f is defined _s =300MHz, carrier frequency 100MHz, bandwidth 20MHz, signal amplitude 20; defining the sampling frequency of BPSK signals as f _s =300MHz, amplitude 20, carrier frequency 50MHz;

3b) The target signal generated in 3a is a label of the training sample obtained by the salich model in step 2 (the saturation output amplitude of the model is 80); wherein the parameters of the saleach model are set as follows _a ＝2，α _φ ＝π/3，β _a ＝1.5625e ^-5 ，β _φ =1; generating a target signal as a training signal of the DNN network, and taking a signal passing through a Saichh model as a label of training data; after nonlinear amplification, the data are obviously changed in both time domain and frequency domain;

3c) The input layer of the DNN network is x (t), the number of nodes of the hidden layer is respectively amplified and reduced, the final number of nodes of the output layer is the same as that of the input layer, and the output signal x after nonlinear amplification is obtained _AP (t); in the experiment, relu (recovered linear velocities) was used as the concealing layerThe hidden layer number of the deep neural network is Num,1024, 2048, \ 8230;, 1024, num; num is the number of points of each sample, and the specific layer number can be adjusted; the loss quadratic cost function in a neural network is:

Loss＝mean(square(x-x _AP )) (10)

the output of each node of the DNN network is a nonlinear activation function to which its inputs are applied; weights between layers in the neural network are optimized through extensive learning, and expected outputs of training samples containing known inputs are learned;

and 4, step 4: inputting a signal generated after the original reference signal passes through a trained network as a new reference signal into an adaptive filter;

the receiver receives signals including a target signal, a noise signal and a self-interference signal, the sum of the signals is a signal before cancellation, and a self-interference signal SI (t) in formula (5) is a target to be cancelled;

4a) Taking LFM signal as an example, the received signal is denoised and sampled to become y (n), which is the cancellation signal input to the adaptive filter and has a duration of 3 μ s,

4b) Delaying the y (n) signal in the step 4a by 0.5 mu s, then carrying out nonlinear processing by using the Sa-Rach model in the step 2 to simulate the nonlinear characteristic of the power amplifier, taking the processed signal as a self-interference signal SI (n), and adding the SI (n), the y (n) and Gaussian white noise to obtain a target cancellation signal;

4c) Inputting the y (n) signal delayed by 0.5 mu s in the step 4b into a DNN neural network as a training sample, taking the signal subjected to nonlinear processing by the Saichh model as a label, and estimating a signal x by a deep neural network _AP (n) as a reference signal for the adaptive filter;

4d) Inputting the signals generated in the steps 4b and 4c into the adaptive filter of the LMS algorithm to obtain an output signal x (n) after cancellation of the adaptive filter, and achieving the result of cancellation of the LFM signal;

the cancellation result can also be achieved by changing the LFM signal in the above steps into a BPSK signal.

Images