Disclosure of Invention
The invention aims to overcome the defects of the prior art, and provides a frequency hopping signal parameter estimation method based on deep learning, which comprises the steps of firstly analyzing the power spectrum characteristics of a frequency hopping signal with a low signal-to-noise ratio under a Gaussian noise background to obtain a frequency set of the frequency hopping signal, and on the basis, constructing a training set of a frequency hopping signal time domain waveform according to the obtained frequency set; and then training the multilayer perceptron to obtain MLP weight matrixes and offset vectors of different frequency signal time domains, and further identifying the signals to be processed to obtain time domain distribution of different frequency signals, so that parameters such as the holding time and the hopping time of the frequency hopping signal are obtained. The time-frequency estimation performance of the method is simulated and analyzed, and the result shows that the method can effectively realize the time-frequency estimation of the frequency hopping signal under the condition of low signal-to-noise ratio.
The purpose of the invention is realized as follows: the invention provides a frequency hopping signal parameter estimation method based on deep learning, which comprises the steps of firstly analyzing the power spectrum characteristics of a frequency hopping signal with a low signal-to-noise ratio under a Gaussian noise background to obtain a frequency set of the frequency hopping signal, and on the basis, constructing a training set of a frequency hopping signal time domain waveform according to the obtained frequency set; and then training the multilayer perceptron to obtain MLP weight matrixes and offset vectors of different frequency signal time domains, and further identifying the signals to be processed to obtain time domain distribution of different frequency signals, so that parameters such as the holding time and the hopping time of the frequency hopping signal are obtained. The time-frequency estimation performance of the method is simulated and analyzed, and the result shows that the method can effectively realize the time-frequency estimation of the frequency hopping signal under the condition of low signal-to-noise ratio.
The method comprises the following specific steps:
step 1: performing power spectrum estimation on the received frequency hopping signal by adopting an AR power spectrum estimation method, and setting a spectrum peak search condition to obtain a frequency set of the frequency hopping signal;
step 2: determining the number of required deep learning networks according to the number of frequencies in the frequency set, constructing a time domain waveform containing noise corresponding to each frequency in the frequency set, and constructing a training set required by the deep learning networks corresponding to each frequency in the frequency set;
and step 3: inputting the training set into each network to obtain a weight matrix and a bias vector of each network, and completing construction of a deep learning network;
and 4, step 4: inputting the received frequency hopping signals into the constructed networks with corresponding frequencies respectively to obtain the output corresponding to each frequency network;
and 5: and smoothing the output of each network to obtain the time domain distribution of the frequency corresponding to the network, and estimating the time-frequency parameters of the received frequency hopping signal.
The invention also includes such structural features:
1. in step 1, an AR power spectrum estimation method is adopted, the order of an AR model is determined through an AIC criterion, and the spectrum peak searching condition is set by using the frequency distribution characteristics of a frequency hopping signal.
2. In step 2, a training set is constructed by using information obtained by receiving signals; assume a frequency in the frequency set of { f1,f2,…,fnConstruction of f1The training set of (a) is:
class 1: time domain signal containing useful signal and noise:
x1(t)=cos(2πf1t)+n1(t);
class 0: time domain signal containing no useful signal but noise:
x0(t)=n0(t);
wherein: n (t) is white Gaussian noise; repeating the construction method for one hundred times to obtain the frequency f1A training set containing one hundred class 1 samples and one hundred class 0 samples of the signal of (a): x is the number of1,100One hundred samples, x, representing class 10,100Represents class 0The one hundred samples of (a); y is1For corresponding to sample X1The classification label of (1); assume a sample length of nsLet n bes=nw×fsWherein n iswIs a frequency of f1The number of cycles of the cosine signal of fsIs the sampling frequency, then X1Is ns×200,Y1Is 1 × 200; other frequency signals in the frequency set are also used for constructing a training set in the same way to obtain n groups of training sets { X) corresponding to n carrier frequency frequencies1,X2,…,XnAnd n sets of labels { Y }1,Y2,…,Yn}。
3. In step 3, a deep neural network is respectively constructed for signals with different frequencies in the frequency set to classify time domain waveforms, the used neural network comprises two hidden layers, the number of nodes of the two hidden layers is 7 and 3 respectively, and a sigmoid function is adopted for an activation function of each layer.
4. In step 4, the received data are respectively input into the constructed neural network, and the length of a sample signal constructed by the training set is assumed to be nsThe length of the signal to be measured passing through the window is nsThen the signals intercepted by the sliding window are respectively input into the trained multilayer perceptron corresponding to n carrier frequencies to obtain n signals with the length of (t.f)s-ns+1) classification set.
5. In step 5, the obtained classification set is subjected to sliding average to eliminate the interference of random noise, and the classification set corresponding to n carrier frequencies obtained after the sliding average is assumed to be { C1,C2,…,CnLet C be assumed1For corresponding carrier frequency of f1Classification set of (1), then C1For corresponding carrier frequency of f1In the time domain, if C1If k points are classified as 1, the frequency of the frequency hopping signal is f1Has a partial hold time of k/fsSecond; classification set C1The point of the first 1 classification in the sequence is the frequency f in the frequency hopping signal1The carrier frequency of (2).
Compared with the prior art, the invention has the beneficial effects that: when the frequency hopping signal frequency set is extracted, the longer time integration is adopted, and compared with a windowing method which needs to give consideration to both time resolution and time resolution in short-time Fourier transform, the obtained frequency hopping signal carrier frequency is more accurate. Meanwhile, the characteristic of strong nonlinear classification performance of the deep neural network is utilized, the network is trained by establishing time domain waveforms of different carrier frequencies of frequency hopping signals, accurate time domain distribution of each carrier frequency can be obtained, frequency hopping signal time-frequency parameters can be estimated under the condition of low signal-to-noise ratio, the limitation of an uncertain principle to a linear time-frequency method can be overcome, the influence of nonlinear time-frequency estimation cross terms does not exist, and high estimation accuracy can be obtained in a time-frequency domain at the same time.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.
Referring to fig. 1, a frequency hopping signal parameter estimation method based on deep learning includes the following steps:
step 1, performing power spectrum estimation on a received frequency hopping signal by adopting an AR power spectrum estimation method, and setting a spectrum peak search condition to obtain a frequency set of the frequency hopping signal;
specifically, an AR power spectrum estimation method is adopted, the order of an AR model is determined through an AIC criterion, and the spectrum peak searching condition is set by using the frequency distribution characteristics of a frequency hopping signal;
step 2, determining the number of required deep learning networks according to the number of frequencies in the frequency set, and constructing a time domain waveform containing noise corresponding to each frequency in the frequency set, so as to construct a training set required by the deep learning networks corresponding to each frequency in the frequency set;
in step 2, a training set is constructed by using information obtained from the received signal. Assume a frequency in the frequency set of { f1,f2,…,fnConstruction of f1The training set of (a) is:
class 1: time-domain signal containing useful signal and noise
x1(t)=cos(2πf1t)+n1(t);
Class 0: time-domain signals containing no useful signal but only noise
x0(t)=n0(t);
Where n (t) is white Gaussian noise. Repeating the construction method for one hundred times to obtain the frequency f1A training set containing one hundred class 1 samples and one hundred class 0 samples of the signal of (a): x is the number of1,100One hundred samples, x, representing class 10,100Representing the one hundred samples of class 0. Y is1For corresponding to sample X1The classification label of (1). Assume a sample length of nsLet n bes=nw×fsWherein n iswIs a frequency of f1The number of cycles of the cosine signal of fsIs the sampling frequency, then X1Is ns×200,Y1Is 1 × 200. Similarly, other frequency signals in the frequency set construct a training set in the same way to obtain n groups of training sets { X) corresponding to n carrier frequency frequencies1,X2,…,XnAnd n sets of labels { Y }1,Y2,…,Yn}。
The training sample length is determined by simulation:
the frequency point of the simulation carrier is {1.4,1.8,2.5,3.5} kHz, the carrier holding time is 50ms, and the take-off time T is0The background noise is white gaussian noise and the SNR is-10 dB.
As can be seen from fig. 2(a), as the frequency increases, the minimum number of waveforms required for the training sample increases, and there is no fixed value for determining the training sample length. Definition k 1000 xnw/fcWherein f iscFor the carrier frequency, by varying the size of the value of kThe change of the sample length is observed, and as can be seen from fig. 2(b), when k is 5, it can be ensured that the mean square error of the retention time of each frequency point is relatively small, and the length of the training sample is as short as possible, thereby reducing the computation amount when the multi-layer perceptron is trained. I.e. the length of the sample in the training set is ns=0.005×fs。
Step 3, inputting the training set into each network to obtain a weight matrix and a bias vector of each network, and completing construction of a deep learning network;
in step 3, a deep neural network is respectively constructed for signals with different frequencies in the frequency set to classify time domain waveforms, the used neural network comprises two hidden layers, the number of nodes of the two hidden layers is 7 and 3 respectively, and a sigmoid function is adopted for an activation function of each layer.
Step 4, inputting the received frequency hopping signals into the constructed networks with corresponding frequencies respectively, thereby obtaining the output corresponding to each frequency network;
in step 4, the received data are respectively input into the constructed neural network, and the length of a sample signal constructed by the training set is assumed to be nsThe length of the signal to be measured passing through the window is nsThen the signals intercepted by the sliding window are respectively input into the trained multilayer perceptron corresponding to n carrier frequencies to obtain n signals with the length of (t.f)s-ns+1) classification set.
And step 5, smoothing the output of each network to obtain the time domain distribution of the frequency corresponding to the network, thereby estimating the time-frequency parameter of the received frequency hopping signal.
In step 5, the obtained classification set is subjected to sliding average to eliminate the interference of random noise, and the classification set corresponding to n carrier frequencies obtained after the sliding average is assumed to be { C1,C2,…,CnLet C be assumed1For corresponding carrier frequency of f1Classification set of (1), then C1For corresponding carrier frequency of f1In the time domain, if C1If k points are classified as 1, the frequency of the frequency hopping signal is f1Has a partial hold time of k/fsAnd second. Classification set C1Of the first class 1The point is that the frequency of the frequency hopping signal is f1The carrier frequency of (2).
The effectiveness of the invention can be verified by the following simulation:
1. the frequency hopping frequency is {1.3,1.7,1.4,1.9,1.8,1.6,1.2,1.5} kHz in sequence, and other simulation parameters are as follows: sampling frequency fs20kHz, 400ms observation time T, and frequency hopping holding time Tk50ms, take-off time T 00. The background noise is white gaussian noise, and the SNR is-10 dB.
FIG. 3 is a comparison between the time-frequency diagram of the STFT and SPWVD, wherein 3(a) is the time-frequency diagram obtained by STFT, and the time-frequency aggregation is poor; 3(b) is a time-frequency diagram obtained by SPWVD, and compared with STFT, the time-frequency aggregation is higher, but more cross item interference is generated; and 3(c) is the method provided by the invention, the time-frequency aggregation is better than that of the two previous traditional methods under the condition of the same signal-to-noise ratio, the time-frequency resolution is higher, and the time-frequency parameters can be estimated more easily.
2. The frequency hopping frequency is {1.3,1.7,1.4,1.9,1.8,1.6,1.2,1.5} kHz in sequence, and other simulation parameters are as follows: sampling frequency fs20kHz, 400ms observation time T, and frequency hopping holding time Tk50ms, take-off time T 00. The background noise is white gaussian noise, and 200 monte carlo experiments are carried out under different signal-to-noise ratios.
Fig. 4(a), fig. 4(b), and fig. 4(c) are diagrams illustrating the comparison between the STFD method and the improved time-frequency ridge line method, wherein the STFD method compares the frequency-hopping frequency retention time, the hopping time, and the estimation accuracy of the frequency-hopping frequency under different snr conditions. It can be seen that the method has higher estimation precision on the frequency hopping frequency holding time, the hopping time and the frequency hopping frequency under different signal-to-noise ratios.
The frequency hopping signal time-frequency parameter is estimated based on the deep learning method, so that the problem that the linear time-frequency method is restricted by an uncertain principle and is not influenced by nonlinear time-frequency estimation cross terms is solved, higher estimation precision can be obtained in the time-frequency domain, and higher time-frequency parameter estimation precision can be obtained under the condition of low signal-to-noise ratio.
In summary, the invention belongs to the technical field of electronic countermeasure and communication, and discloses a frequency hopping signal parameter estimation method based on deep learning, which comprises the following steps: 1) carrying out power spectrum estimation on the received frequency hopping signal to obtain a frequency set of the frequency hopping signal; 2) Determining the number of required deep learning networks according to the number of frequencies in the frequency set, and constructing a training set required by the deep learning networks corresponding to each frequency in the frequency set; 3) inputting the training set into each network to complete the construction of the deep learning network; 4) inputting the received signals into the constructed networks with corresponding frequencies respectively, thereby obtaining the output of the network corresponding to each frequency; 5) And performing smoothing processing on the output of each network to estimate the time-frequency parameters of the received frequency hopping signal. The method has higher estimation precision on the time-frequency parameters of the frequency hopping signal under the condition of low signal-to-noise ratio, and has important significance on the processing of the frequency hopping signal.