A spread spectrum signal detection method integrating cognitive enhanced compressed sensing and artificial neural network

Abstract

The present invention belongs to the field of spread spectrum communication signal detection. The detection of spread spectrum communication signals is an important part of modern military and civilian electronic countermeasures. The ultra-wideband and ultra-high frequency of modern spread spectrum technology bring a burden to the detection technology based on traditional Nyquist sampling. At the same time, the existing spread spectrum signal detection algorithm still has shortcomings in terms of adaptability and online processing capabilities. In order to break through the above bottleneck, the present invention integrates cognitive enhanced compressed sensing and artificial neural network ideas, and discloses an adaptive compression detection method for spread spectrum communication signals. On the one hand, the idea of ​​cognitive enhanced compressed sensing can reduce the signal sampling rate and processing complexity; on the other hand, the artificial neural network can enhance the algorithm's signal feature extraction ability and online processing ability. Based on the above two points, the present invention can detect spread spectrum communication signals online, adaptively and accurately at a lower signal sampling rate in the absence of prior knowledge of the signal.

Description

Technical Field

[0001] The present invention belongs to the field of spread spectrum communication signal detection, and specifically is an adaptive spread spectrum communication signal detection method integrating artificial neural network and cognitive enhanced compressed sensing. It involves artificial neural network, adaptive compressed sensing, signal detection, signal posterior probability analysis, and spread spectrum technology. Background Art

[0002] Spread spectrum communication has strong resistance to human interference, narrowband interference, and multipath interference, and was first used in military communication systems. In recent years, the theory and application of spread spectrum communication technology have developed very rapidly, and have been widely used in various fields of military and civilian communications, thus promoting the rapid development of the communication industry. The detection of spread spectrum signals is a prerequisite for subsequent signal parameter estimation and signal recovery.

[0003] Frequency hopping, direct sequence spread spectrum and time hopping are the most commonly used spread spectrum technologies. Time hopping is the selection of signal jumps in transmission time, so as to achieve the expansion of signal frequency domain. Time hopping needs to be used simultaneously with frequency hopping or direct sequence spread spectrum. Frequency hopping signals use carrier frequency hopping to make the signal carrier switch quickly within a wide frequency band according to a certain pseudo-random sequence, so as to achieve the expansion of signal frequency domain. Direct sequence spread spectrum signals use pseudo-random sequence modulation to expand the signal to a frequency band much wider than the baseband. According to different spread spectrum methods, the academic community has proposed a variety of signal detection methods.

[0004] (1) Spread spectrum signal detection method based on traditional Nyquist sampling

[0005] Traditional signal detection methods are based on Nyquist sampling, mainly including methods based on time-frequency analysis and methods based on the maximum likelihood criterion. With the development of data analysis theory, domestic and foreign scholars have proposed a series of improvement schemes for methods based on time-frequency analysis. For example, Kyung-Gyu Lee et al. combined the time-frequency characteristics and statistical characteristics of the signal to improve the accuracy of the spread spectrum signal parameter estimation method. Zhaozhao Zhang et al. proposed a spread spectrum signal detection method that does not require prior knowledge of the signal based on the statistical characteristics analysis of the second-order mean of the autocorrelation function.

[0006] The existing spread spectrum signal detection method based on Nyquist sampling has good detection capabilities, but the requirements for ultra-wideband, ultra-high frequency, high bit rate, anti-interference and other performance indicators of military spread spectrum technology in actual combat are getting higher and higher, which brings a burden to the detection technology based on traditional Nyquist sampling: the increase in spread spectrum bandwidth means the continuous increase in sampling rate, and the increase in sampling rate increases the hardware cost of the signal system. However, when the bandwidth reaches a certain width, such methods will face hardware bottlenecks; the increase in signal sampling frequency means that the system needs to process more signal samples per unit time, which brings a great burden to the processing system.

[0007] (2) Spread spectrum signal detection method based on compressed sensing

[0008] The research and application of compressed sensing has pointed out a new direction for the development of spread spectrum signal interception technology. The compressed sensing theory shows that if a signal can be sparsely expressed based on a dictionary or in a linear transformation domain, then in theory, the signal can be measured at a sampling rate far lower than the Nyquist frequency, and the signal can be reconstructed by a certain method without causing signal and information loss.

[0009] Recently, domestic and foreign scholars have conducted some research on compressed sensing methods for spread spectrum signals. In the compressed sensing measurement phase, Chen Ying and Shengheng Liu et al. studied the design method of compressed sensing matrix based on orthogonal matrix and signal prior fuzzy function, respectively, and improved the quality of compressed sensing information recovery under low signal-to-noise ratio conditions. In the sparse expression dictionary construction phase, Jin Yan et al. and Maja et al. constructed a dictionary based on the parameterized model of frequency hopping signals; Peng Li et al. and Shuang Wu et al. studied the construction of a sparse expression dictionary for signals based on a parameterized direct sequence spread spectrum signal model. In the information extraction stage, most academics first reconstruct the signal from the compressed sensing data, and then analyze and obtain the required signal detection results from the reconstructed signal. In view of the fact that the support set of the spread spectrum code of the direct sequence spread spectrum signal is large and the signal detection and estimation methods are complex, Cheng Yanhe et al. studied a two-stage reconstruction method, which reduced the complexity of the method to a certain extent.

[0010] The existing spread spectrum detection method based on compressed sensing can reduce the sampling rate of the signal to a certain extent, but it still needs further improvement in the following aspects:

[0011] ① Existing spread spectrum signal detection based on compressed sensing cannot perform online adaptive design of compressed sensing matrix based on the a posteriori cognition of signal and channel. Therefore, it cannot be guaranteed that such methods can obtain the best signal detection performance with the least measurement samples.

[0012] ② Most of the existing spread spectrum signal detection based on compressed sensing first requires the reconstruction of compressed sensing data. However, the complexity of compressed sensing reconstruction is relatively high, so it is difficult to process high-rate spread spectrum communication signals online.

[0013] The present invention is based on the fusion of cognitive enhanced compressed sensing and artificial neural network theory. It can efficiently obtain the posterior information of signal distribution without reconstructing the signal, and use artificial neural network to quickly design the measurement kernel to be used in the next step of compressed sensing, thereby realizing the efficient, fast and adaptive detection of spread spectrum signals in space with a small number of measurements. Summary of the invention

[0014] The present invention aims to solve the technical problems that the bandwidth of spread spectrum communication signals is large, detection is difficult using traditional non-cooperative sampling methods, and the existing compressed sampling signal detection technology has poor adaptability.

[0015] The specific technical solution of the present invention to solve the above problem is: a spread spectrum signal detection method integrating cognitive enhanced compressed sensing and artificial neural network, the steps of which are:

[0016] (1) The received channel input is first preprocessed by a bandpass filter, and the preprocessed result is measured using the initialized compressed sensing kernel to obtain certain compressed sensing data;

[0017] (2) using a certain segment of compressed sensing data that has been obtained, to obtain the a posteriori distribution probability of the discretized parameters of the spread spectrum signal to be detected (for example, for frequency hopping signals, the a posteriori usage probability of each frequency band is used; for direct sequence spread spectrum signals, the a posteriori usage probability of each spreading code is used; for time hopping, the a posteriori usage probability of each time slot is used);

[0018] (3) The estimated parameter posterior distribution probability is input into the trained artificial neural network in the form of a vector, and then the compressed sensing measurement kernel (i.e., the non-zero sub-blocks in the compressed measurement matrix, see Figure 2 );

[0019] (4) preprocessing the latest received channel input by a bandpass filter, and continuing to measure the preprocessed result using the compressed sensing kernel obtained in step (3) to further obtain compressed sensing data;

[0020] (5) Stop or repeat steps (2) to (4) several times, and the system can collect a certain amount of compressed sensing data;

[0021] (6) Take a certain section of measurement data, calculate the data energy and judge by the energy threshold, that is, it can be determined whether the spread spectrum signal exists.

[0022] In the present invention, the artificial neural network used for adaptive compressed sensing kernel design is trained by stochastic gradient descent method: the training data is a randomly generated parameter distribution, and the penalty function of the neural network training is the inverse of the mutual information of the system received signal and the compressed sensing data obtained by the compressed sensing kernel output by the neural network under the assumption of the current parameter posterior probability distribution. That is, the compressed sensing kernel designed by the artificial neural network in the present invention maximizes the mutual information between the parameterized modeled signal in the space and the compressed sensing data under the assumption of the current parameter posterior probability distribution, thereby realizing the gradual increase of the signal-to-noise ratio in the compressed sensing data, and finally obtaining the improvement of the detection accuracy of the spread spectrum signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is the system framework structure of the present invention;

[0024] Figure 2 is a compressed measurement matrix structure used for signal detection (the figure takes the compressed measurement matrix structure used for five measurements as an example, but the number of measurements is not limited to five and can be taken arbitrarily);

[0025] Figure 3 It is a schematic diagram of an artificial neural network used for the design of a compressed sensing measurement kernel in the present invention;

[0026] Figure 4 It is a schematic diagram of a process of performing spread spectrum signal detection using the present invention;

[0027] Figure 5 Taking a frequency hopping signal as an example, the performance of the present invention is compared with that of a compression detection using a random measurement kernel;

[0028] Figure 6 Taking a direct sequence spread spectrum signal as an example, the performance of the present invention is compared with that of a compression detection using a random measurement kernel. DETAILED DESCRIPTION

[0029] The system framework structure of the present invention is as follows Figure 1 The specific design scheme can be described in the following two parts.

[0030] (1) Design of cognitive enhanced compressed sensing spread spectrum signal detection system

[0031] exist Figure 1 In the described system, the compressed sensing matrix is ​​a block diagonal matrix (such as Figure 2 As shown in FIG. 1 , each matrix sub-block represents a measurement kernel required for a compressed sensing measurement, and the ratio of the length to the width of the compressed sensing matrix is ​​the compression ratio of the compressed sensing. The present invention combines the a posteriori analysis of the signal to adaptively design the compressed sensing measurement kernel, and detects the spread spectrum signal based on the detection of the compressed sensing data energy.

[0032] The cognitive enhancement of compressed sensing for signal detection is to design the measurement kernel of each compressed sensing measurement line by line in a cognitive adaptive manner. Based on a certain segment of compressed sensing measurement data that has been obtained, the present invention converts the conditional mutual information I(x m ;y m |φ s ,φ 2 ,...,φ m-1 ,y s ,y 2 ,...,y m-1 ) is defined as the task-specific information of signal detection, where 1≤s≤m-1, φ i refers to the measurement kernel used in the i-th compressed sensing; x i With y i Respectively in φ i The channel output and compressed sensing data during the measurement period are represented by column vectors and satisfy:

[0033] y i =φ i x i (1)

[0034] The task-specific information can be obtained through a posteriori probability analysis of key signal parameters (e.g., for frequency hopping signals, the a posteriori usage probability of each frequency band; for direct sequence spread spectrum signals, the a posteriori usage probability of each spread spectrum code; for time hopping, the a posteriori usage probability of each time slot). In theory, the design of the compressed sensing kernel coefficients of the present invention is mainly based on solving the optimization problem in formula (2), thereby realizing cognitive enhancement and self-adaptation of compressed sensing.

[0035]

[0036] In order to make the energy of the data obtained from each compression measurement comparable, the coefficients of each row of the compressed sensing kernel obtained by formula (2) can be Modulo unitization, we can get the mth compressed sensing measurement kernel φ m .

[0037] Using the newly designed compressed sensing kernel, the system will further perform compressed sensing measurement on the signal, update the a posteriori information, and design the compressed sensing measurement kernel. The system flow chart is as follows: Figure 4 shown.

[0038] In the present invention, the signal existence detection is completed by observing the energy of the compressed sensing data. k ,y k+1 ,...,y M , then the energy of this segment of compressed sensing data can be obtained by Indicates that is the 1-2 norm of the vector. In the signal detection stage, an energy threshold T is set. If the energy If it is greater than or equal to T, it is determined that the signal exists; otherwise, it is determined that the signal does not exist.

[0039] (2) Non-cooperative adaptive detection method for spread spectrum signals integrating cognitive enhanced compressed sensing and artificial neural network

[0040] Theoretically, the solution to formula (2) can be performed by gradient recursive loop optimization. However, the implementation of loop optimization consumes a lot of time and computing resources. In order to maximize the efficiency of online adaptive design of compressed sensing measurement kernel, the present invention integrates cognitive enhanced compressed sensing and artificial neural network to realize efficient adaptive detection method of spread spectrum signal. This part mainly includes three main points:

[0041] ① Use a discrete signal parameter ξ to describe the signal characteristics (for example, for a frequency hopping signal, the spectrum can be divided equally and ξ is taken as each sub-frequency band; for a direct sequence spread spectrum signal, ξ can be taken as each spreading code; for a time hopping signal, the signal transmission period can be divided evenly and ξ is taken as each sub-time slot).

[0042] ② Offline training of artificial neural network:

[0043] The artificial neural network used in step ② is a fully connected neural network, such as Figure 3 As shown in the figure, the input layer of the network is the posterior probability distribution value of the discretized parameter, and the output is the compressed sensing measurement kernel coefficient obtained by adaptive design. In order to make the energy of the data obtained by each compression measurement comparable, the coefficients of each row of the compressed sensing kernel can be normalized by 1-2 to obtain the mth compressed sensing measurement kernel.

[0044] The present invention adopts random discretized parameter distribution as training data of artificial neural network, takes the opposite number of specific task information as penalty function, and trains the artificial neural network based on back propagation method.

[0045] ③After the artificial neural network training is completed, the compressed sensing measurement kernel is designed online using the artificial neural network.

[0046] This enables efficient adaptive compressed sensing and accurate signal detection of the system.

[0047] The present invention can efficiently and accurately detect spread spectrum signals at a relatively low sampling rate. Compared with the commonly used signal compression detection method based on random compressed sensing measurement kernel, the measurement accuracy of the present invention is significantly improved. For example, Figure 5 and Figure 6The performance simulation results of using a random measurement kernel and the measurement kernel of the present invention in a frequency hopping signal and a direct sequence spread spectrum signal are shown respectively.

[0048] Figure 5 The frequency hopping signal used in the experiment is a Gaussian frequency shift keying signal, and the carrier frequency jumps between 2.402GHz and 2.480GHz, with a frequency hopping carrier interval of 1MHz. The symbol period is 1μs, the frequency hopping period is 625μs, and the total measurement time is 80μs. Each sub-block of the compressed sensing measurement matrix is ​​set to 1 row. When a random compressed sensing measurement kernel is used, the measurement kernel is generated by an independent Gaussian random variable with unit variance and zero mean, and each row of the measurement matrix is ​​normalized by 1-2 modulo.

[0049] When the system false alarm rate is 0.01, Figure 5 They are respectively the missed detection rate curves for 320 and 640 measurements. When the measurement number is 320, the spectrum from 2.4005 GHz to 2.4805 GHz is divided into 20 parts equally, and ξ is taken as the occupancy of each sub-band; when the measurement number is 640, the spectrum from 2.4005 GHz to 2.4805 GHz is divided into 10 parts equally, and ξ is taken as the occupancy of each sub-band. Corresponding to the above different spectrum segmentation methods, the number of nodes in the input layer of the neural network is 20 and 10 respectively, and five hidden layers are taken, and the number of layers is 128, 256, 512, 256, and 128 respectively. The number of nodes in the output layer of the neural network depends on the compression ratio of compressed sensing.

[0050] Depend on Figure 5 It can be seen that in order to achieve the same signal leakage rate performance, when the sampling number is 320 times, the present invention can save about 7dB of signal-to-noise ratio compared with the case of using the random compressed sensing kernel. When the sampling number is 640 times, the present invention can save about 5dB of signal-to-noise ratio compared with the case of using the random compressed sensing kernel.

[0051] Figure 6 The direct sequence spread spectrum signal used in the method is a binary phase shift keying signal. The spread spectrum code takes the maximum length sequence of order 1 to 5, with a total of 234 spread spectrum codes, and the longest chip length can reach 31. The symbol period is 1μs, and the total measurement time is 40μs. Each sub-block of the compressed sensing measurement matrix is ​​set to 1 row. When a random compressed sensing measurement kernel is used, the measurement kernel is generated by an independent unit variance zero mean Gaussian random variable, and each row of the measurement matrix is ​​normalized by 1-2 modulo.

[0052] Using the present invention, ξ can be taken as the use of each spread spectrum code. The number of nodes in the input layer of the neural network is 234, and three hidden layers are taken, whose number of layers are 512, 350, and 256 respectively. The number of nodes in the output layer of the neural network is set to the total length of the measurement kernel in one code element period: 62. After each neural network outputs the measurement kernel coefficient, the corresponding coefficient is taken from it to design the current measurement kernel. When the system false alarm rate is 0.01, Figure 6 These are the missed detection rate curves for 240, 360, and 480 measurements respectively. Figure 6 It can be seen that in order to achieve the same signal leakage rate performance, when the sampling number is 240 times, the present invention can save about 3dB of signal-to-noise ratio compared with the case of using the random compressed sensing kernel. When the sampling number is 360 times, the present invention can save about 2dB of signal-to-noise ratio compared with the case of using the random compressed sensing kernel. When the sampling number is 480 times, the present invention can save about 1dB of signal-to-noise ratio compared with the case of using the random compressed sensing kernel.

Claims

1. A non-cooperative detection method for spread spectrum communication signals integrating cognitive enhanced compressed sensing and deep learning, characterized in that: (1) The received channel input is first preprocessed by a bandpass filter, and the preprocessed result is multiplied by the compressed sensing kernel initialized in the compressed measurement matrix, which is a non-zero sub-block in the compressed sensing matrix, to obtain the initial measurement data; (2) updating the posterior probability of the parameter distribution according to the existing measurement data vector and the prior probability of the signal parameter distribution; (3) The posterior probability is imported into the pre-trained artificial neural network, and the output coefficients are organized into the form of a compressed sensing measurement kernel. In order to make the energy of the data obtained in each compression measurement comparable, each row can be normalized by 1-2 modulo to obtain the compressed sensing measurement kernel required for the next compression measurement; (4) The newly received channel input is preprocessed by a bandpass filter, and the preprocessed result is multiplied by the measurement kernel just designed to obtain new partial measurement data; (5) stopping or repeating the steps (1) to (5) several times to obtain a final compressed sensing measurement data vector; (6) Based on the measurement data vector obtained above, the measurement data energy is calculated, and the presence or absence of a signal is judged based on the threshold. If the data energy is less than the threshold, it is judged that the signal does not exist; if the data energy is greater than the threshold, it is judged that the signal exists.

2. The structure and training method of the artificial neural network in the non-cooperative detection of spread spectrum communication signals integrating cognitive enhanced compressed sensing and deep learning as claimed in claim 1, characterized in that: (1) The artificial neural network uses a fully connected layer. The artificial neural network consists of an input layer, a hidden layer, and an output layer; (2) The input layer is the posterior probability distribution value of the discretized signal parameters; (3) The output layer is the coefficients required to design the compressed sensing matrix measurement kernel. These coefficients are arranged in the matrix form of the measurement kernel. In order to make the energy of the data obtained from each compressed measurement comparable, each row can be normalized by 1-2 modules to obtain the latest designed compressed sensing measurement kernel. (4) Using random discretized signal parameter distribution as training data for artificial neural networks; (5) The inverse of the specific task information is used as the penalty function, that is, on the premise of obtaining the probability of the posterior distribution of the parameters, the inverse of the mutual information between the signal and the measurement data is used as the penalty function to train the artificial neural network.

Images