Rec signal detection method based on otfs modulation and deep learning

Abstract

The application discloses a REC signal detection method based on OTFS modulation and deep learning, and the implementation steps are as follows: OTFS demodulation is performed on a REC signal; a target state information in the REC signal is detected by using an approximate message passing algorithm of sparse Bayesian learning; a channel is sequentially subjected to equalization operation and preprocessing by using the state information; a conditional generative adversarial network constructed is trained by using the preprocessed received signal; and the preprocessed received signal is sent into the trained conditional generative adversarial network to perform communication decision, so that a recovered REC communication signal transmitted by a UAV label of our side is obtained. The approximate message passing algorithm of sparse Bayesian learning is used to reduce the time complexity and space complexity of an algorithm in a target detection process, and the special properties of a channel matrix are used to reduce the inverse operation in an equalization process, so that the calculation complexity is further reduced, and the system operation efficiency is improved. The deep learning model is introduced into the communication receiver design, and the communication reliability can still be maintained when a high-speed moving target is used.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, and more specifically relates to a radar-embedded communication (REC) signal detection method based on orthogonal time-frequency space (OTFS) modulation and deep learning within the field of radar communication technology. This invention can be used to detect communication signals in radar-embedded communication in high-speed mobile environments. Background Technology

[0002] Radar-embedded communication technology has emerged as a popular research field in recent years, driven by the scarcity of spectrum resources and the development of new communication technologies. With the rise of the information age, wireless communication systems face increasingly complex electromagnetic environments and limited spectrum resources. Traditional communication systems, due to their inherent openness and broadcast nature, are vulnerable to eavesdropping and interception of unencrypted communication content, posing a potential threat to communication security. Another significant drawback is electromagnetic interference. Wireless communication networks are typically susceptible to electromagnetic interference from enemy radar signals, affecting the continuity and reliability of communication. Enemy radar may use electronic jamming to render the receiver's communication system inoperable, severely impacting command and control. Therefore, more covert and anti-jamming communication methods are needed to ensure the security of communication systems. Radar-embedded communication remodulates individual radar pulses into communication waveforms, embedding them in the radar echo for transmission, thus ensuring communication covertness while increasing communication speed.

[0003] The National University of Defense Technology of the Chinese People's Liberation Army disclosed a radar embedded communication method based on Single Value Decomposition (SVD) in its patent application, "A Radar Embedded Communication Waveform Design Method Based on Singular Value Decomposition" (Patent Application No.: 202210786867, Publication No.: CN 115201759 A). This method first models the radar embedded communication and its channel, then overcomes the limitations of eigenvalue decomposition-based radar embedded communication waveform design by using SVD to extract features of the echo matrix, thereby constructing the radar embedded communication waveform. Compared to traditional radar embedded communication waveform design methods, this method offers more reliable communication performance and a lower probability of interception. However, this method still has shortcomings. Because the channel modeling is based on an ideal Gaussian white noise channel, in practical applications, radar embedded communication systems are often used to detect and track high-speed moving targets. In this scenario, communication will experience Doppler frequency shift and multipath fading, and the channel will deteriorate into a time-varying frequency-selective bicolor dispersive channel, leading to a sharp decrease in communication reliability.

[0004] Xi'an University of Electronic Science and Technology disclosed a radar embedded communication system using OTFS waveforms as radar signals in its patent application "Radar Embedded Communication Method Based on OTFS Modulation" (Patent Application No.: 202310111332.6, Publication No.: CN 116125457 A). This system utilizes a matched filter (MF) in the time-delay-Doppler domain for channel estimation, transforming the time-varying frequency-selective bidispersive wireless channel into the time-delay-Doppler domain. This ensures that all symbols in the transmission unit experience almost identical and slowly changing sparse channels, thereby obtaining full channel diversity in both time and frequency. This method overcomes the problem in existing technologies where high-speed movement of the target causes severe inter-carrier interference due to Doppler frequency shift, leading to a decrease in radar detection accuracy. Although this invention uses OTFS waveforms as radar signals to suppress Doppler frequency shift caused by high-speed movement, the system still has shortcomings. The radar signal processing still employs a matched filtering algorithm, and the equalization algorithm involves numerous matrix inversion operations, whose high complexity limits efficiency in practical applications. Moreover, the communication receiver is still based on a mature radar-embedded communication solution, which uses a lot of Monte Carlo simulation and consumes a lot of computing resources. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the existing technology by proposing a REC signal detection method based on OTFS modulation and deep learning. The OTFS technology solves the problem of excessive computational complexity and the need for a large number of Monte Carlo simulations in the traditional REC signal detection process in high-speed mobile environments.

[0006] The approach to achieving the objectives of this invention is to address the high complexity of radar detection by employing a parameter estimation algorithm based on Approximate Message Passing (AMP). This algorithm simplifies the radar target signal model by utilizing prior information such as the empirical range of parameter values ​​and the sparse structure of the radar channel, resulting in a sparse signal estimation model. Furthermore, by combining the approximate message passing method with the traditional Sparse Bayesian Learning (SBL) algorithm, the time and space complexity of the algorithm can be significantly reduced in the detection of friendly UAV targets. Simultaneously, the special properties of the channel matrix are utilized to reduce the inversion operations during the equalization process, further reducing computational complexity. This significantly improves the system's operating efficiency. Given the need for extensive Monte Carlo simulations in traditional communication receiver simulations, this invention introduces a deep learning model into the design of the communication receiver. A Conditional Generative Adversarial Network (CGAN) is used to learn the true mapping between the received signal and the signal transmitted by the friendly UAV, enabling the discrimination of signals received by the cooperative receiver. By relying on a pre-trained model for online deployment, the received signal is demodulated in real time, thereby restoring the embedded communication signal transmitted by the tag of the friendly drone.

[0007] The specific steps for implementing this invention include the following:

[0008] Step 1: Demodulate the received REC signal using OTFS;

[0009] Step 2: Detect the target state information of our UAV in the REC signal using an approximate message-passing algorithm based on sparse Bayesian learning;

[0010] Step 3: Perform equalization on the channel using the state information;

[0011] Step 4: Preprocess the equalized received signal;

[0012] Step 5: Construct a conditional generative adversarial network;

[0013] Step 6: Train the conditional generative adversarial network using the preprocessed received signal;

[0014] Step 7: Using the same processing method as in Step 4, preprocess the received signal to be detected, and send the preprocessed received signal into the trained conditional generative adversarial network for communication decision-making to obtain the recovered embedded communication signal transmitted by our UAV tag.

[0015] Compared with existing technologies, the present invention has the following advantages:

[0016] First, this invention uses OTFS waveforms as radar signals to construct a radar-embedded communication system, overcoming the problem of decreased communication reliability in high-speed mobile scenarios in existing technologies. In radar processing, an approximate message-passing parameter estimation algorithm is employed, combining it with the traditional sparse Bayesian learning algorithm, which significantly reduces the time and space complexity of the algorithm during target detection. In equalization, a low-complexity equalization algorithm based on eigenma factorization is used, leveraging the special properties of the channel matrix to reduce inversion operations during the equalization process, further lowering computational complexity. This significantly improves the system's operating efficiency.

[0017] Secondly, this invention introduces a deep learning model into the design of the communication receiver. By using a conditional generative adversarial network (GAN) to learn the true mapping between the received signal and the signal transmitted by the friendly UAV, it overcomes the deficiency of traditional communication receiver simulations which require extensive Monte Carlo simulations. This allows for the discrimination of signals received by the cooperative receiver. Compared to traditional DLD receivers, the proposed CGAN receiver has significant advantages in design and implementation difficulty, as well as communication performance, and maintains good communication reliability even when facing high-speed moving targets. Attached Figure Description

[0018] Figure 1 This is a flowchart of the present invention;

[0019] Figure 2 This is a flowchart of the approximate message passing algorithm for sparse Bayesian learning in this invention;

[0020] Figure 3 This is a block diagram of the radar embedded communication receiver of the CGAN of the present invention;

[0021] Figure 4 This is a block diagram of the CGAN network structure of the present invention;

[0022] Figure 5 This is a simulation diagram of the present invention, wherein, Figure 5 (a) is a graph showing the peak sidelobe ratio of the SBL-AMP algorithm and the MF algorithm; Figure 5 (b) is a graph showing the symbol error rate of the CGAN receiver and the DLD receiver under different main space sizes according to the present invention; Figure 5 (c) is a graph showing the symbol error rate of the CGAN receiver at different speeds according to the present invention. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0024] Reference Figure 1 The implementation steps of the embodiments of the present invention will be described in further detail below.

[0025] Step 1: Demodulate the received REC signal using OTFS.

[0026] The expression for the REC signal is as follows:

[0027] r2(t)=α k c k (t)+r1(t)*p(t)+n(t)

[0028] Where r2(t) represents the REC signal received by the cooperative receiver at time t, and c k (t) represents the k-th embedded communication waveform at time t, α k c represents the time t. k The power constraint factor of (t), r1(t) represents the radar signal received by the tag on our UAV at time t, p(t) represents the environmental scattering at time t, * represents the convolution operation, and n(t) represents the system noise at time t.

[0029] The steps for OTFS demodulation are as follows:

[0030] The first step is to process the received REC signal using the Wigner transform.

[0031] The second step is to sample the transformed signal to recover the time-frequency signal.

[0032] The third step involves using the Inverse Symplectic Finite Fourier Transform (ISFFT) to convert the time-frequency signal back to the time-delay-Doppler domain symbol y[k,l], thus completing the entire reception and demodulation process.

[0033] Step 2: Use the approximate message passing algorithm based on sparse Bayesian learning to detect the target status information of our UAV in the REC signal.

[0034] Reference Figure 2 The approximate message passing algorithm flow of sparse Bayesian learning in this embodiment of the invention will be described in further detail.

[0035] The steps for detecting the target state information of our UAV in the REC signal using the approximate message-passing algorithm based on sparse Bayesian learning are as follows:

[0036] The first step is to convert the demodulated REC signal into the following vector form based on the input-output relationship of the OTFS signal:

[0037] y = H DD x+w

[0038] Where y, x, w represent the vector forms of y[k,l], x[k,l], and w[k,l], respectively. The (k+Nl)th element in y, x, w is equal to y[k,l], x[k,l], and w[k,l], respectively, where x[k,l] represents the transmitted signal of the system, and w[k,l] represents the variance of σ. 2 Gaussian white noise, k and l represent the Doppler tap and the time delay tap, respectively, k = 0, ..., N-1, l = 0, ..., M-1, N and M represent the number of subcarriers and the number of symbols, respectively, and their values ​​are all integers greater than 0; H DD The channel matrix in the time-delay-Doppler domain is represented.

[0039] The second step is to transform the above vector expression into the following form:

[0040]

[0041] Where h represents the vector form of h[k,l], The (k+Nl)th element of h is h[k,l], where h[k,l] represents the complex gain at the target location of our UAV to be detected. This represents the transformed equivalent channel matrix. The values ​​in the first column are equal to the values ​​in x, and the values ​​in the remaining columns are cyclically shifted according to the following formula:

[0042]

[0043]

[0044] in, Representing the equivalent channel matrix The element values ​​located at coordinates (m, n) are given, where k1 and k2 represent Doppler taps, l1 and l2 represent time delay taps, k1 = 0, ..., N-1, l1 = 0, ..., M-1, k2 = 0, ..., N-1, l2 = 0, ..., M-1, n = 0, ..., N-1, m = 0, ..., M-1, e (·) This represents an exponential function with the natural exponent e as the base, j represents the imaginary unit, and π represents pi.

[0045] This method draws inspiration from the idea of ​​communication symbol detection, and is to... It is equivalent to the known channel matrix, while the channel vector h is equivalent to the symbol to be detected.

[0046] The third step is to utilize the maximum target detection range R set in the radar system. max and maximum detection speed Vmax Pruning the channel vector h transforms the target's time delay taps and Doppler taps into:

[0047]

[0048]

[0049] Among them, M e and N e These represent the time delay tap and Doppler tap after pruning, respectively, where B represents the system bandwidth, and f... c This indicates the center carrier frequency of the transmitter.

[0050] At this point, the dimension of the channel vector h decreases from NM×1 to N e M e ×1, correspondingly, matrix The dimension is reduced from NM×NM to NM×N e M e .

[0051] The fourth step involves extracting the model of the received signal y to obtain a sparse model, as follows:

[0052]

[0053] in, This represents the extracted receive vector. G represents the number of samples, and A represents the observation matrix after sampling at the corresponding position of y. This represents the observation noise extracted at the corresponding position of y.

[0054] The fifth step is to solve for the signal to be estimated in the sparse model using the approximate message passing algorithm described below.

[0055] h t+1 =η τ (h t +A * z t )

[0056]

[0057] Among them, h t+1 ,h t ,h t-1 Let η represent the estimated values ​​of the signal h to be estimated after the (t+1), (t), and (t-1)th iterations, respectively. τ (·) denotes the sampling filter operator for the t-th iteration, η τ (h t + A *z t ) = sign(h t +A * z t )((h t +A * z t )-τ) + The sign function is represented by `sign(·)`, the subscript `+` indicates that only positive values ​​are considered, and `τ` represents the threshold parameter, derived from the formula `τ = bτ0`, where `b > 0`, and `τ0` is the initial value (h). t +A * z t The root mean square error of z, where the superscript * indicates the transpose operation. t Let z represent the residual value during the t-th iteration. t-1 η represents the residual value during the (t-1)th iteration. τ ′(h t-1 +A * z t-1 ) represents η τ (h t +A * z t The first derivative of ) This indicates that the average value of each element in vector u is calculated. u = η τ ′(h t-1 +A * z t-1 ), where δ represents the compression ratio during the extraction process.

[0058] The sixth step is to use the channel vector estimated by the above model. The value is transformed into h[k,l]. By analyzing the distribution of h[k,l] values, the time delay and Doppler taps corresponding to its peak values ​​are identified. These taps correspond to the relative time delay τ and Doppler frequency shift ν between our UAV target and the radar signal transmitter, respectively. The channel state information of the system at the time delay τ and Doppler frequency shift ν is then h(τ,ν). The detected distance and velocity of our UAV target are as follows:

[0059]

[0060]

[0061] Where R and V represent the relative distance and relative speed between our UAV target and the radar transmitter, respectively, and c0 represents the speed of light.

[0062] Step 3: Use the state information to perform equalization on the channel.

[0063] The steps for performing channel equalization using state information are as follows:

[0064] The first step is to transform the time-delay-Doppler channel state information h(τ,ν) into a time-domain channel matrix according to the following formula:

[0065]

[0066]

[0067] Among them, H DD Represents the time-delay-Doppler domain channel matrix. H represents the time-domain channel matrix. [·] M Represents the modulo-M operation, F N Represents the N-point discrete Fourier transform matrix. The superscript H indicates the conjugate transpose operation, and the superscript -1 indicates the matrix inversion operation.

[0068] The second step is to decompose the channel matrix H using the following formula:

[0069]

[0070]

[0071] where F M represents the M-point discrete Fourier transform matrix, D represents a diagonal matrix whose diagonal elements are equal to the eigenvalues of the corresponding positions in the H matrix, diag(·) represents a matrix with the vector as the main diagonal element value, and vec(·) represents a matrix vectorization operation, H1 represents the first column of the matrix H.

[0072] Thirdly, the equalized received signal expression is obtained according to the following formula:

[0073]

[0074] wherein, represents the equalized received signal, and I represents a unit matrix,

[0075] After equalization, the estimated can be sent to the receiver for decision to detect the communication performance of the system. Since can be realized by two-dimensional fast Fourier transform and its inverse transform, and is a diagonal matrix, when the number of data symbols in the OTFS frame is large, compared with directly using matrix inversion, the calculation complexity can be significantly reduced.

[0076] With reference to Figure 3 , the signal preprocessing and CGAN demodulation process of the embodiment of the present application are further described in detail.

[0077] Step 4, preprocessing the equalized received signal.

[0078] The preprocessing steps are as follows:

[0079] Firstly, a filter set is constructed according to the following formula:

[0080] w k = (RR H + λ max I NM ) -1 c k

[0081] wherein w k represents the kth filter in the filter set, k = 1, 2, …, K, and the value of K is equal to the size of the communication waveform set, R represents the Toeplitz matrix obtained by cyclically shifting the signal r1(t) received by our tag, and λ max c represents the largest eigenvalue of a non-principal subspace. k This represents the k-th REC communication waveform.

[0082] The second step is to send the equalized received signal to the filter set and extract the data features of the received signal according to the following formula:

[0083]

[0084] Among them, z k x(n) represents the sample set x(n) after passing through the k-th filter, x(n) = [x n x n-1 ... x n-NM+1 ] T n >> NM, x n Indicates the signal after equalization The remaining values ​​after sampling are x. n Circular shift.

[0085] The third step is to perform a modulo operation on the data features of each extracted received signal to obtain the feature set to be decided.

[0086] Step 5: Construct a conditional generative adversarial network.

[0087] Reference Figure 4 The following is a further detailed description of the CGAN network structure block diagram of an embodiment of the present invention.

[0088] The conditional generative adversarial network includes a generator and a discriminator. The generator is structured as follows: an input layer, a first convolutional layer, a second convolutional layer, a first pooling layer, a third convolutional layer, a second pooling layer, a fourth convolutional layer, a third pooling layer, a nonlinear activation layer, an adaptive max-pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer are sequentially connected in series. Each pooling layer is connected by a batch normalization operation and a ReLU activation function. The input layer is set to have 4×256 channels, and the kernel sizes of the first to fourth convolutional layers are set to 7×1, 3×1, 3×1, and 3×1, respectively, with the number of kernels set to 6. 4, 128, 256, 512, with a stride of 1; the first to third pooling layers are all max pooling layers, each with a size of 1×2; the first to third fully connected layers reduce the dimension of the input data after adaptive max pooling from 512 to 128, 64, and 5 respectively; the output layer further reduces the dimension of the input data from 5 to 1; the discriminator is composed of a first fully connected layer and a second fully connected layer connected in series, which respectively increase the dimension of the discriminator's input information from 2 to 128 and then reduce it to 1; the last fully connected layer of the discriminator uses the Sigmoid activation function to output the discrimination result, and the remaining fully connected layers are connected to the ReLU activation function.

[0089] Reference Figure 4 The training and decision-making process of the conditional generative adversarial network in the embodiments of the present invention will be described in further detail.

[0090] Step 6: Train the conditional generative adversarial network using the preprocessed received signal.

[0091] The process of training a conditional generative adversarial network using the preprocessed received signal is as follows: the preprocessed received signal and constraints are input into the conditional generative adversarial network, and the weight values ​​of the network parameters are iteratively updated using an adaptive gradient algorithm until the total loss function of the network converges, thus obtaining the trained conditional generative adversarial network.

[0092] During network training, the set of received signals |z| after feature extraction via filters is... k (n)| is denoted as Z, the sampled version of the channel coefficients h[k,l] h(n) is denoted as Φ, and the signal recovered after training is denoted as The actual transmitted signal is represented by X.

[0093] Figure 4 The network consists of two parts. During the offline training phase, the generator is responsible for recovering the transmitted signal based on the conditional input, and the discriminator is responsible for recovering the transmitted signal based on the actual transmitted signal X and the generator's output. The discriminator identifies the authenticity of data and performs inverse optimization of the generator's generation performance using a loss function. Once the model is nearing convergence and training is complete, the model or training weights can be saved for online deployment. The generator's input consists of two parts: the received signal Z after feature extraction and the conditional input Φ; the discriminator's input consists of three parts: the generator's output... The actual transmitted signal X and the conditional input Φ.

[0094] The constraints are the relative time delay between our UAV target and the radar signal transmitter, and the channel state information at the Doppler shift.

[0095] The total loss function of the network is:

[0096]

[0097] in, The function represents the expected value, log(·) represents the logarithmic operation with the natural constant e as the base, D θ (X,Φ) represents the discriminator's judgment result on the actual transmitted signal, D θ (G(Z,Φ))) represents the discrimination result of the discriminator on the signal to be detected, X represents the actual transmitted signal, Z represents the set of received signals after preprocessing, and Φ represents the channel state information after sampling.

[0098] Step 7: Using the same processing method as in Step 4, preprocess the received signal to be detected, and send the preprocessed received signal into the trained conditional generative adversarial network for communication decision-making to obtain the recovered embedded communication signal transmitted by our UAV tag.

[0099] The technical effects of the present invention will be further illustrated by the following simulation experiments.

[0100] 1. Simulation experimental conditions:

[0101] The hardware platform for the simulation experiment of this invention is as follows: the processor is an 11th Gen Intel(R) Core with a processor frequency of 3.50GHz, and the graphics card is an NVIDIA GeForce RTX 3090.

[0102] The software platform for the simulation experiment of this invention is Windows 10 system and Matlab R2022b.

[0103] This document describes how to build a conditional generative adversarial network (GAN) using PyTorch with Python 3.7. The PyTorch version is 1.7.1+cu110.

[0104] 2. Simulation Experiment Content and Result Analysis:

[0105] The simulation experiment of this invention uses the radar embedded communication method based on OTFS modulation, and simulates the peak sidelobe ratio and communication symbol error rate under the time-delay-Doppler channel. The results are as follows: Figure 5 As shown.

[0106] The method of this invention uses 16 OTFS radar signal modulation carriers and 16 symbols, with a radar transmission center carrier frequency of 78 GHz, a bandwidth of 2 MHz, a subcarrier spacing of 128 kHz, 4QAM modulation, a main space size of 128, and an embedded communication waveform set size of 4.

[0107] In the conditional generative adversarial network constructed in this invention, the integrated received signal after filtering is used as the training set. The dimension of the input signal is 100000×4×256. The test set is generated in the same way as the training set, with a dimension of 104000×4×256. That is, 40000 sets of data are prepared for testing at each signal-to-noise ratio. The network is trained using parameters of BatchSize of 128, number of iterations of 200, and initial learning rate of 2e-5.

[0108] To verify the simulation effect of the present invention, the computational complexity of the present invention and the traditional matched filtering algorithm were obtained, as shown in Table 1.

[0109] Table 1: Comparison of Computational Complexity

[0110]

[0111] As shown in Table 1, the object detection algorithm of this invention has a complexity of . The computational complexity of the equalization algorithm after eigenvalue decomposition is... The overall complexity is Compared to traditional matched filtering algorithms, the object detection algorithm complexity, the equalization algorithm complexity without feature matrix decomposition, and the overall complexity are all [missing information]. The computational complexity of this invention is significantly reduced, thereby significantly reducing the demand for computing resources.

[0112] To verify the effectiveness of the simulation experiment of this invention, the peak sidelobe ratio was obtained under different signal-to-noise ratios, and the relationship between the obtained peak sidelobe ratio and signal-to-noise ratio was plotted as follows: Figure 5 (a) shows the three curves. Figure 5 In (a), the horizontal axis represents the signal-to-noise ratio of the radar-embedded communication system, in dB, and the vertical axis represents the peak-to-sidelobe ratio. Figure 5 The curve marked with a square in (a) represents the relationship between the peak sidelobe ratio and the signal-to-noise ratio obtained by simulation after sequential extraction using the method proposed in this invention. Figure 5 The curve marked with an asterisk in (a) represents the relationship between the peak sidelobe ratio and the signal-to-noise ratio obtained by simulation after random sampling using the method proposed in this invention. Figure 5 The curve marked with a circle in (a) represents the relationship between peak sidelobe ratio and signal-to-noise ratio obtained by simulation using the traditional matched filtering algorithm.

[0113] By varying the size of the main space and the speed of the friendly UAV, the symbol error rate (BER) under different signal-to-noise ratios (SNRs) was obtained. The relationship between the obtained BER and SNR was then plotted as follows: Figure 5 (b) and Figure 5 The curve shown in (c) Figure 5 In (b), the horizontal axis represents the signal-to-noise ratio of the radar-embedded communication system, and the vertical axis represents the symbol error rate. Figure 5 (b) The curve marked with a triangle represents the relationship between the symbol error rate and the signal-to-noise ratio obtained by simulation using existing technology with a main space of 128. Figure 5 (b) The curve marked with a square represents the relationship between the symbol error rate and the signal-to-noise ratio obtained by simulation using existing technology with a main space of 192. Figure 5 (b) The curve marked with an asterisk represents the relationship between the symbol error rate and the signal-to-noise ratio obtained by simulation using the method proposed in this invention with a main space of 128. Figure 5 (b) The curve marked with a circle represents the relationship between the symbol error rate and the signal-to-noise ratio obtained by simulation using the method proposed in this invention with a principal space of 192. Figure 5 (c) The curves marked with an asterisk represent the relationship between symbol error rate and signal-to-noise ratio (SNR) obtained by simulation under relatively static conditions using the method proposed in this invention. The curves marked with triangles represent the relationship between symbol error rate and SNR obtained by simulation under low-speed conditions using the method proposed in this invention. The curves marked with circles represent the relationship between symbol error rate and SNR obtained by simulation under high-speed conditions using the method proposed in this invention. The low-speed condition indicates that the distance between the friendly UAV target and the radar transmitter is 150m and the speed is 30.05m / s. The high-speed condition indicates that the distance between the friendly UAV target and the radar transmitter is 900m and the speed is 195.31m / s.

[0114] In the simulation experiment of this invention, the prior art used refers to the radar embedded communication method implemented by singular value decomposition disclosed by the National University of Defense Technology of the Chinese People's Liberation Army in its patent application document "A Radar Embedded Communication Waveform Design Method Based on Singular Value Decomposition" (patent application number: 202210786867, application publication number: CN 115201759 A).

[0115] The following is a simulation. Figure 5 The effects of the present invention will be further described.

[0116] from Figure 5 As shown in (a), the pruned sequential decimation algorithm achieves a higher peak-to-side-lobe ratio (PSLR) than the traditional matched filtering algorithm. However, the pruned random decimation algorithm exhibits a larger variation in its PSLR estimate. Specifically, the performance of the random decimation algorithm is significantly affected by the signal-to-noise ratio (SNR), showing a wider range of variation, while the performance of the sequential decimation algorithm remains relatively stable under different SNRs. This indicates that, within a certain range, the sequential decimation algorithm is relatively more robust.

[0117] observe Figure 5 The results in (b) show that the symbol error rate (BER) performance of the system significantly improves with the increase of the main space size. Compared with the existing DLD receiver, the CGAN receiver proposed in this invention exhibits superior BER performance at all main space sizes. When the main space value is 128, the CGAN receiver's BER performance improves by 1–2 dB; while when the main space value is 196, the performance improvement is particularly significant, approximately 2–3 dB. Therefore, it can be clearly concluded that the CGAN receiver proposed in this invention outperforms the traditional DLD receiver in terms of communication performance.

[0118] contrast Figure 5 The simulation results in (c) lead to the following conclusion: Although the training performance of the CGAN network may vary slightly at different speeds, its impact on the symbol error rate can be considered negligible overall. This result verifies that the CGAN network has strong robustness and stability in dealing with radar signal transmission at different speeds. Although speed variations may have some impact on network training, the overall performance fluctuation is not significant, which provides strong support for the application of the CGAN network in high-speed moving scenarios.

[0119] In summary, this invention discloses a REC signal detection method based on OTFS modulation and deep learning. It uses OTFS waveforms as radar signals to construct a radar-embedded communication system, addressing the problem of decreased communication reliability in high-speed moving scenarios. In radar processing, an approximate message-passing algorithm combining sparse Bayesian learning and a low-complexity equalization algorithm are employed, significantly reducing the time and space complexity of the algorithm and thus significantly improving the system's operating efficiency. By introducing a deep learning model into the design of the communication receiver and learning the true mapping between the received signal and the signal transmitted by a friendly UAV through a conditional generative adversarial network, the problem of requiring extensive Monte Carlo simulations in traditional communication receiver simulations is solved. Simulation results show that compared to traditional DLD receivers, the proposed CGAN receiver has significant advantages in design and implementation difficulty, as well as communication performance, and maintains good communication reliability even when facing high-speed moving targets.

Claims

1. A method for detecting a REC signal based on OTFS modulation and deep learning, characterized in that, The OTFS waveform is used as a radar signal to construct a radar embedded communication system, a conditional generative adversarial network is constructed, and communication decision is made in the trained conditional generative adversarial network; the specific steps of the method include the following: Step 1, OTFS demodulation is performed on the received radar embedded communication REC signal; Step 2, the target state information of the unmanned aerial vehicle of the user is detected in the REC signal by using the approximate message passing algorithm of sparse Bayesian learning; the steps of the algorithm are as follows: First, the demodulated REC signal is converted into the following vector form according to the input-output relationship of the OTFS signal: ； wherein, respectively represent , and in vector form, , the value of the th element in , and , represents a transmit signal of the system, represents a Gaussian white noise with variance , , and respectively represent a Doppler tap and a delay tap, , respectively represent the number of subcarriers and the number of symbols, and both take integer values greater than 0; represents a channel matrix in the delay-Doppler domain, ; Second, the above vector expression is equivalent to the following form: ； in, express The vector form, , The The elements are , This indicates the complex gain at the target location of our drone to be detected. This represents the transformed equivalent channel matrix. , The first column of values ​​and If the element values ​​in the first column are equal, the values ​​in the remaining columns are cyclically shifted according to the following formula: ； ； wherein denotes the equivalent channel matrix located in at coordinates and denotes a Doppler tap, and denotes a delay tap, , , , denotes the exponential function with natural exponent , denotes the imaginary unit, denotes the circle constant; Third step, the maximum detection distance of the target is set in the radar system and the maximum detection speed Pruning the channel vector , the time delay tap and the Doppler tap of the target are transformed into: ； wherein, and denote the pruned delay taps and Doppler taps, respectively, denotes the bandwidth of the system, denotes the center carrier frequency of the transmitter; In a fourth step, the received signal is extracted from the model to obtain a sparse form of the model as follows: ； wherein, represents the received vector after decimation, , represents the number of decimation, represents the observation matrix after decimation at the corresponding position of , , represents the observation noise after decimation at the corresponding position of , ; Fifth, the approximate message passing algorithm is used to solve the to-be-estimated signal in the sparse model: ； ； wherein respectively denote the signal to be estimated the estimate after the first iteration, denotes the sample filter operator for the first iteration, , denotes the sign function, the subscript denotes the positive sign only, denotes the threshold parameter, which is derived from the equation , , is the initial root mean square error, the superscript denotes the transpose operation, denotes the residual value during the first iteration process, denotes the residual value during the first iteration process, denotes the first derivative, denotes the average of each element in the calculation vector , , , denotes the compression rate during the decimation process, ; Step 6: The channel vector estimated by the above model is converted to ; by analyzing the distribution of the values, the peak value corresponding to the time delay and Doppler taps are identified, which correspond to the relative time delay and Doppler frequency shift between the UAV target and the radar signal transmitter , respectively , then the channel state information of the system at the time delay and Doppler frequency shift is , and the distance and speed of the detected UAV target are respectively: ； wherein, and R and V represent the relative distance and relative velocity between the UAV target and the radar transmitter, respectively, c represents the speed of light; Step 3, the channel is equalized by using the state information: In a first step, the delay-Doppler channel state information is converted into a time-domain channel matrix according to ； ； wherein denotes a delay-Doppler domain channel matrix, , denotes a time domain channel matrix, , denotes a modulo operation, denotes an N-point discrete Fourier transform matrix, denotes a Kronecker product, the upper index denotes a conjugate transpose operation, the upper index denotes a matrix inversion operation; In a second step, the channel matrix H is decomposed using the following equation: H = U * D * V ； ； wherein, denotes the M-point Discrete Fourier Transform matrix, denotes a diagonal matrix whose diagonal elements have values equal to the eigenvalues of the matrix in the corresponding positions, , denotes a matrix whose main diagonal elements have values equal to the vector denotes the vectorization operation of a matrix, , denotes the first column of the matrix ; Third, the equalized received signal expression is obtained according to the following formula: ； wherein denotes the equalized received signal, denotes the unit matrix, ; Step 4, the equalized received signal is preprocessed: First, a filter set is constructed according to the following formula: ； in, Represents the first in the filter set One filter, , , The value of is equal to the size of the communication waveform set. , This indicates the signal received by our tag. The Toplitz matrix obtained after cyclic shifting Represents the largest eigenvalue of a non-principal subspace. Indicates the first One REC communication waveform; Second, the equalized received signal is sent to the filter set, and the data features of the received signal are extracted according to the following formula: ； wherein, represents the set of samples after passing through the represents the equalized signal the result after sampling, the remaining values being cyclic shifts of .​​​​ Third, the data features of each received signal are subjected to a modulus operation to obtain a feature set to be judged; Step 5, a conditional generative adversarial network is constructed; Step 6, the conditional generative adversarial network is trained by using the preprocessed received signal; Step 7, the same processing method as step 4 is used to preprocess the to-be-detected received signal, the preprocessed received signal is sent into the trained conditional generative adversarial network for communication decision, and the recovered embedded communication signal transmitted by the unmanned aerial vehicle of the user is obtained.

2. The method of claim 1, wherein, The expression of the REC signal in step 1 is as follows: ； wherein, denotes the REC signal received by the cooperating receiver at the time instant, denotes the time instant the embedded communication waveform, denotes the time instant power constraint factor, , denotes the radar signal received by the tag on our drone at the time instant, denotes the environmental scattering at the time instant, denotes the convolution operation, denotes the system noise at the time instant.

3. The method of claim 1, wherein, The steps of the OTFS demodulation in step 1 are as follows: First, the received REC signal is processed by using the Wigner transform; Second, the transformed signal is sampled to restore the time-frequency signal; In the third step, the time-frequency signal is converted back to the time-delay-Doppler domain symbol by using the finite Fourier transform of sine , and the whole receiving demodulation process is completed.

4. The method of claim 1, wherein, The conditional generative adversarial network under the conditions described in step 5 comprises a generator and a discriminator; the structure of the generator is sequentially connected by an input layer, a first convolutional layer, a second convolutional layer, a first pooling layer, a third convolutional layer, a second pooling layer, a fourth convolutional layer, a third pooling layer, a nonlinear activation layer, an adaptive max-pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer and an output layer; each pooling layer is connected by a batch normalization operation and a ReLU activation function; the input layer is set to 4x256 channels, the sizes of the convolution kernels of the first to fourth convolutional layers are set to 7x1, 3x1, 3x1 and 3x1 respectively, the numbers of the convolution kernels are set to 64, 128, 256 and 512 respectively, and the strides are all set to 1; the first to third pooling layers are all max-pooling layers with a size of 1x2; the first to third fully connected layers reduce the dimensions of the input data after adaptive max-pooling from 512 to 128, 64 and 5 respectively; the output layer reduces the dimension of the input data from 5 to 1; the discriminator is sequentially connected by a first fully connected layer and a second fully connected layer, which respectively increase the dimensions of the input information of the discriminator from 2 to 128 and then reduce them to 1; the last fully connected layer of the discriminator uses a Sigmoid activation function to output a discrimination result, and the other fully connected layers are connected with a ReLU activation function.

5. The method of claim 1, wherein, The training process of the conditional generative adversarial network using the preprocessed received signal in step 6 is: inputting the preprocessed received signal and the constraint condition into the conditional generative adversarial network, using an adaptive gradient algorithm to iteratively update the weight values of the network parameters until the total loss function of the network converges, and obtaining the trained conditional generative adversarial network.

6. The method of claim 5, wherein, The constraint condition is the channel state information at the relative time delay and Doppler shift between the unmanned aerial vehicle target and the radar signal transmitter.

7. The method of claim 5, wherein, The total loss function of the network is: ； wherein, denotes a function for taking an expectation value, denotes a logarithm operation with a natural constant as a base, denotes a discrimination result of the discriminator for a real transmission signal, denotes a discrimination result of the discriminator for a signal to be detected, denotes a real transmission signal, denotes a set of pre-processed reception signals, denotes a sampled channel state information.

Images