Radar signal RF covert communication waveform recognition method based on convolutional neural network

Abstract

The present invention relates to the field of radar waveform modulation recognition. To solve the problem of identifying covert waveforms, the present invention provides a method for identifying radar signal radio frequency covert communication waveforms based on a convolutional neural network. The steps are as follows: Step 1: For unknown signals intercepted by a radar reconnaissance aircraft, complete the detection of the LFM waveform and parameter estimation including the initial frequency f0 and the frequency modulation slope μ; Step 2: Generate two types of waveforms including this covert waveform and the original LFM waveform to construct a data set; Step 3: Obtain the received waveform at the receiving end. Step 4: Complete the construction of the waveform data set; Step 5: Build a convolutional neural network; Step 6: Use mini-batch optimization for network training, and use the SGD optimizer as the optimizer. The present invention is mainly used in radar waveform modulation recognition scenarios.

Description

Technical Field

[0001] The present invention relates to the field of radar waveform modulation recognition, and in particular to a method for recognizing radar signal radio frequency covert communication waveforms based on a convolutional neural network. Background Art

[0002] In modern electronic warfare, in order to reduce the probability of detection of one's own radar waveform, many researchers have designed radio frequency waveforms with low detection probability. For example, researchers [1] designed radio frequency waveforms with low detection probability based on spread spectrum technology, whose power spectrum density is less than the noise power threshold. Researchers [2] designed signals that are very similar to noise signals, such as chaotic signals, to achieve covert information transmission. The above two waveforms have achieved a certain concealment effect against some traditional detection methods such as energy detection. However, when faced with the new detection method based on the cyclostationary analysis theory proposed by WAGardner [3] in the late 1970s, the above two waveforms cannot provide a guarantee of low detection probability.

[0003] Linear Frequency Modulation (LFM) chirp signals are sinusoidal signals whose instantaneous frequency varies linearly with time. They have a large spectrum bandwidth and good autocorrelation. They are often used in radar and sonar systems. Due to their strong anti-interference ability, some scholars have also studied the application of LFM chirp signals in communication systems. In recent years, Zhiping Zhang et al. used LFM chirp radar signals as carriers and improved the traditional BPSK modulation method. They proposed a new modulation framework consisting of binary reduced phase shift keying (BRPSK) and variable symbol width, and proposed a new radio frequency covert communication waveform based on LFM chirp radar signals. Under noise-free conditions, the detection characteristics of this waveform and the original chirp signal in all existing transform domain methods, including time domain, frequency domain, and cyclic domain, are almost identical [4]. At present, there is no work to solve the problem of identifying the above special covert waveforms.

[0004] In recent years, deep learning, with its data-driven nature, lack of prior feature design, and end-to-end training, has been increasingly applied to problems in communication systems, such as automatic modulation recognition for adaptive demodulators[5] and individual identification of radiators[6]. These applications abstract communication problems into classification or regression tasks in deep learning, thereby utilizing deep network models to perform inference tasks. For some of the more complex problems in the communication field, constructing feature engineering based on prior knowledge is difficult, and the end-to-end learning concept in deep learning provides a new approach to solving complex problems.

[0005] [1]Schoolcraft R.Low probability of detection communications-LPDwaveform design and detection techniques[C] / / IEEE Military CommunicationsConference.IEEE,1991.

[0006] [2]Lau F, Tse C K. Chaos-based digital communication systems: operating principles, analysis methods, and performance evaluation[M]. Springer, 2003.

[0007] [3]Gardner W A.The spectral correlation theory of cyclostationarytime-series-ScienceDirect[J].Signal Processing,1986,11(1):13-36.

[0008] [4]Zhang Z,Nowak MJ,Qu Y,et al.RF Steganography via LFM Chirp RadarSignals[J].IEEE Transactions on Aerospace&Electronic Systems,2017:1-1.

[0009] [5]Xue Yuhong. Research on communication signal modulation recognition technology based on deep learning[D]. Harbin Institute of Technology, 2020.

[0010] [6] Niu Haonan, Wang Wencan, Liu Qingbo. Radar emitter identification based on convolutional neural network[J]. Modern Defense Technology, 2021, 49(03): 130-136. Summary of the Invention

[0011] To overcome the shortcomings of the existing technology, the present invention aims to solve the problem of identifying covert waveforms. The waveform recognition problem is processed into a binary classification problem of the original LFM waveform and the covert waveform. Then, a covert waveform recognition algorithm based on a convolutional neural network is proposed. To this end, the technical solution adopted by the present invention is a method for recognizing radar signal RF covert communication waveforms based on a convolutional neural network. The steps are as follows:

[0012] Step 1: For the unknown signal intercepted by the radar reconnaissance aircraft, complete the detection of the LFM waveform and the estimation of parameters including the initial frequency f0 and the frequency modulation slope μ;

[0013] Step 2: Generate two types of waveforms including this hidden waveform and the original LFM waveform to construct a data set;

[0014] Step 3: After the transmitter completes the generation of the transmission waveform, the channel condition adopts an additive white Gaussian noise channel. The waveform transmitted by the transmitter passes through the additive white Gaussian noise channel and the received waveform is obtained at the receiver.

[0015] Step 4: Under different signal-to-noise ratio conditions, randomly generate information bits and modulation phase sequences to generate different hidden waveform samples and original LFM signal samples. Use the same sampling rate to generate waveform data and attach binary labels to the two types of waveform data to complete the construction of the waveform dataset.

[0016] Step 5: Build a convolutional neural network;

[0017] Step 6: The network is trained using mini-batch optimization, and the optimizer is SGD optimizer.

[0018] The specific process of step 2 is that the pseudo-random phase generator generates n binary data {b1,...,b i-1 ,b i ,...,b n}, where b i = ±1, generating a pseudo-random phase sequence {φ1,...,φ i-1 ,φ i ,...,φ n}, set the initial code element width Then according to formula (1) and the pseudo-random modulation phase sequence {φ1,...,φ i-1 ,φ i ,...,φ n}, calculate the code element width sequence

[0019]

[0020] The pulse shaping filter generates a baseband waveform based on the symbol width sequence and data polarity, and then the binary phase modulator modulates the baseband waveform based on the generated phase sequence and data polarity. The multiplier then multiplies the signal s(t) generated by the LFM waveform generator and takes the real part of the signal to complete the RF covert waveform. The production of As shown in formula (2), where Ac is the signal amplitude, φ i is the modulation phase corresponding to each bit, φ I is the initial phase of the signal:

[0021]

[0022] The convolutional neural network has a total of 7 layers, including 2 convolutional layers, 2 maximum pooling layers, and 3 fully connected layers. The convolution kernel size of the first convolutional layer is 3*3, the number of convolution kernels is 20, the stride is 1, and there is no padding; the second layer is a maximum pooling layer, the convolution kernel size is 2*2, the stride is 1; the third layer is a convolutional layer, the convolution kernel size is 3*3, the number of convolution kernels is 50, the stride is 1, and there is no padding; the fourth layer is a maximum pooling layer, the convolution kernel size is 2*2, and the stride is 1. The last three layers of the network are fully connected layers, with 550, 200, and 2 neurons respectively; the last layer uses softmax as its activation function to output the probability value of the sample belonging to one of the two categories, and the activation function of the two convolutional layers both use sigmoid as their activation function;

[0023] First, the input signal is preprocessed, and the signal with a length of 500 is transformed into a 50*10 two-dimensional matrix. Then it enters the convolution layer to extract the features of the signal data. The extracted features are further reduced in dimension through the pooling layer to extract key information. The features are then combined through the fully connected layer. Finally, the softmax layer outputs the probability value of the signal sample belonging to the two categories of hidden waveform and original waveform.

[0024] The characteristics and beneficial effects of the present invention are:

[0025] The hidden waveform recognition algorithm based on convolutional neural network proposed in this invention is unable to complete the recognition of the hidden waveform by combining the characteristics of deep learning that does not require prior feature design and autonomous learning of data structure information. It can realize the recognition of the new hidden waveform, and when the signal-to-noise ratio is greater than -2dB, the recognition accuracy can reach more than 97%, proving the effectiveness of the hidden waveform recognizer based on convolutional neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The figure shows the overall block diagram of the process of generating the RF covert communication waveform based on the LFM chirp radar signal.

[0027] Figure 2 Schematic diagram of the original LFM waveform, BPSK-modulated LFM waveform and this new covert waveform.

[0028] Figure 3 This is a structural diagram of the convolutional neural network proposed in this invention.

[0029] Figure 4 The results of network recognition accuracy under different signal-to-noise ratios are shown in the figure. DETAILED DESCRIPTION

[0030] The present invention first simplifies the problem of identifying this hidden waveform into a simple two-classification problem. By utilizing the characteristics of neural networks that do not require prior feature design and are data-driven, a hidden waveform recognition algorithm based on convolutional neural networks is proposed to complete the recognition of this hidden waveform.

[0031] The present invention is applied to the field of radar waveform modulation recognition, which is a new type of recognition problem of radio frequency covert communication waveform based on LFM chirp radar signal, and proposes a waveform recognition algorithm based on convolutional neural network.

[0032] The technical solution adopted by the present invention is a hidden waveform detection method based on a convolutional neural network, the steps of which are as follows:

[0033] Step 1: After completing the LFM waveform detection and parameter estimation (including the initial frequency f0 and the frequency modulation slope μ) for the unknown signal intercepted by the radar reconnaissance aircraft, the method of the present invention is required to further identify the new radio frequency covert communication waveform based on the LFM chirp radar signal and the original LFM waveform, because traditional detection methods cannot further identify this new radio frequency covert communication waveform based on the LFM chirp radar signal and the original LFM waveform.

[0034] Step 2: Since the present invention adopts a waveform recognition method based on a data-driven neural network, we need to generate two types of waveforms including this hidden waveform and the original LFM waveform to construct a data set. Figure 1 The overall block diagram of the generation process of the radio frequency covert communication waveform based on the LFM chirp radar signal is as follows: the pseudo-random phase generator generates the n binary data {b1,...,b i-1 ,b i ,...,b n}, where b i = ±1, generating a pseudo-random phase sequence {φ1,...,φ i-1 ,φ i ,...,φ n Unlike the traditional BPSK modulation phase angle with a phase difference of π, the modulation phase values used in the phase sequence are generally small and the difference between them is small. Set the initial symbol width Then according to formula (1) and the pseudo-random modulation phase sequence {φ1,...,φ i-1 ,φ i ,...,φ n}, calculate the code element width sequence

[0035]

[0036] The pulse shaping filter generates a baseband waveform based on the symbol width sequence and data polarity, and then the binary phase modulator modulates the baseband waveform based on the generated phase sequence and data polarity. The multiplier then multiplies the signal s(t) generated by the LFM waveform generator and takes the real part of the signal to complete the RF covert waveform. The production of. As shown in formula (2), where A c is the signal amplitude, φ i is the modulation phase corresponding to each bit, φ I is the initial phase of the signal. Figure 2 The figure shows the original LFM waveform, the BPSK-modulated LFM waveform and the schematic diagram of this new concealed waveform. It can be seen that under noise-free conditions, the time domain waveforms of this concealed waveform and the original LFM are very similar, almost exactly the same.

[0037]

[0038] Step 3: After the transmitter completes the generation of the transmission waveform, the channel condition adopts an additive white Gaussian noise channel. The waveform transmitted by the transmitter passes through the additive white Gaussian noise channel and the received waveform is obtained at the receiver.

[0039] Step 4: Under different signal-to-noise ratio conditions, randomly generate information bits and modulation phase sequences to generate different hidden waveform samples and original LFM signal samples. Use the same sampling rate to generate waveform data, and attach binary labels to both types of waveform data to complete the construction of the waveform dataset. The experimental parameters used in this invention are shown in Table 1.

[0040] Table 1 Experimental parameter settings

[0041]

[0042] The modulation phase adopted by the present invention is a discrete random variable with a value of 12.5 to 17.5° and an interval of 0.1°, with a total of 51 discrete phases. An information bit of length 9 is randomly generated, thereby generating a random phase sequence of length 9, setting the initial code element width to 1us, thereby generating a code element width sequence of length 9, which is truncated or completed so that the total length of the code element width sequence is 10us. The sampling frequency is set to 50MHz, and the sampling time and pulse width are 10us, so the length of a single received signal sample generated is 500. An AWGN channel environment is adopted, and the signal-to-noise ratio is set to -10 to 10dB with an interval of 2, so there are a total of 10 signal-to-noise ratios. 10,200 concealed waveform samples and 10,200 original LFM waveform samples are generated under each signal-to-noise ratio. The training set and the test set are divided into 7:3. The training set is a mixture of samples generated under all signal-to-noise ratios, so the training set contains a total of 142,800 samples.

[0043] Step 5: Build the convolutional neural network proposed in this invention. The network structure diagram is as follows: Figure 3 As shown in the figure, the network has a total of 7 layers, including 2 convolutional layers, 2 max pooling layers, and 3 fully connected layers. The first convolutional layer has a kernel size of 3*3, 20 kernels, a stride of 1, and no padding. The second layer is a max pooling layer with a kernel size of 2*2 and a stride of 1. The third layer is a convolutional layer with a kernel size of 3*3, 50 kernels, a stride of 1, and no padding. The fourth layer is a max pooling layer with a kernel size of 2*2 and a stride of 1. The last three layers are fully connected layers, with 550, 200, and 2 neurons, respectively. The last layer uses softmax as its activation function to output the probability that the sample belongs to one of the two categories. Both convolutional layers use sigmoid as their activation function. First, the input signal is preprocessed, and the signal with a length of 500 is transformed into a 50*10 two-dimensional matrix. Then it enters the convolution layer to extract the features of the signal data. The extracted features are further reduced in dimension through the pooling layer to extract key information. Then, the features are combined through the fully connected layer. Finally, the softmax layer outputs the probability value of the signal sample belonging to the two categories (hidden waveform and original waveform).

[0044] Step 6: The network is trained using mini-batch optimization, with a batch size of 256 and a training epoch of 100. The SGD optimizer is used, with an initial learning rate of 0.1 and a learning rate decay strategy that decays to 0.5 of the original value every 10 epochs. After training, the network is tested at each signal-to-noise ratio. The test results are as follows: Figure 4 As shown in the figure, when the signal-to-noise ratio is greater than -2dB, the accuracy can reach more than 97%, proving the effectiveness of the hidden waveform recognizer based on convolutional neural network.

[0045] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with this technical field within the technical scope disclosed in the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for recognizing radio frequency covert communication waveforms of radar signals based on convolutional neural networks, characterized in that the steps as follows: Step 1: For the unknown signal intercepted by the radar reconnaissance aircraft, complete the detection of the LFM waveform and the estimation of parameters including the initial frequency f0 and the frequency modulation slope μ; Step 2: Generate two types of waveforms including hidden waveforms and original LFM waveforms to construct a data set; Step 3: After the transmitter completes the generation of the transmit waveform, the channel condition adopts an additive white Gaussian noise channel. The waveform transmitted by the transmitter passes through the additive white Gaussian noise channel and is then received at the receiver. Step 4: Under different signal-to-noise ratio conditions, randomly generate information bits and modulation phase sequences to generate different hidden waveform samples and original LFM signal samples. Use the same sampling rate to generate waveform data and attach binary labels to the two types of waveform data to complete the construction of the waveform dataset. Step 5: Build a convolutional neural network, use it to process the waveform data, and output the probability values of the two categories: hidden waveform and original waveform; Step 6: The convolutional neural network is trained using mini-batch optimization, and the SGD optimizer is used.

2. The method for recognizing radar signal radio frequency covert communication waveform based on convolutional neural network according to claim 1, characterized in that: The specific process of step 2 is that the pseudo-random phase generator generates n binary data {b1,...,b i-1 ,b i ,...,b n }, where b i = ±1, generating a pseudo-random phase sequence {φ1,...,φ i-1 ,φ i ,...,φ n }, set the initial code element width Then according to formula (1) and the pseudo-random modulation phase sequence {φ1,...,φ i-1 ,φ i ,...,φ n }, calculate the code element width sequence The pulse shaping filter generates a baseband waveform based on the symbol width sequence and data polarity, and then the binary phase modulator modulates the baseband waveform based on the generated phase sequence and data polarity. The multiplier then multiplies the signal s(t) generated by the LFM waveform generator and takes the real part of the signal to complete the RF covert waveform. The production of As shown in formula (2), where A c is the signal amplitude, φ i is the modulation phase corresponding to each bit, φ I is the initial phase of the signal:

3. The method for recognizing radar signal radio frequency covert communication waveform based on convolutional neural network according to claim 1, characterized in that: The convolutional neural network has a total of 7 layers, including 2 convolutional layers, 2 maximum pooling layers, and 3 fully connected layers. The convolution kernel size of the first convolutional layer of the network is 3*3, the number of convolution kernels is 20, the step size is 1, and there is no padding; the second layer is the maximum pooling layer, the convolution kernel size is 2*2, the step size is 1; the third layer is the convolution layer, the convolution kernel size is 3*3, the number of convolution kernels is 50, the step size is 1, and there is no padding; the fourth layer is the maximum pooling layer, the convolution kernel size is 2*2, and the step size is 1; the last three layers of the network are fully connected layers, and the number of neurons is 550, 200, and 2 respectively; the last layer uses softmax as its activation function to output the probability value of the sample being one of the two categories, and the activation function of the two convolutional layers both use sigmoid as their activation function; First, the input signal is preprocessed, and the signal with a length of 500 is transformed into a 50*10 two-dimensional matrix. Then it enters the convolution layer to extract the features of the signal data. The extracted features are further reduced in dimension through the pooling layer to extract key information. The features are then combined through the fully connected layer. Finally, the softmax layer outputs the probability value of the signal sample belonging to the two categories of hidden waveform and original waveform.

Images