CN114978381B - Compressed sensing processing method and device for broadband signals based on deep learning - Google Patents

Abstract

The invention discloses a compressed sensing processing method and device of broadband signals based on deep learning, comprising the following steps: performing multiple set sampling on an input broadband signal in a time domain to obtain a discrete sampling sequence; preprocessing the obtained discrete sampling sequence and the input broadband signal and constructing a data set; designing a broadband signal reconstruction neural network model ADMM-net based on deep learning, and initializing the ADMM-net; in the model training stage, inputting signal data of a training set into the ADMM-net, and continuously minimizing a loss function through an optimization algorithm to obtain optimal neural network parameters; and inputting the signal data of the test set into the trained ADMM-net model to obtain a reconstructed broadband signal. The ADMM-net combined with the deep learning in the invention has higher signal reconstruction accuracy under the conditions of lower sampling rate and lower signal-to-noise ratio after being trained by a network model, and has certain generalization capability.

Description

Compressed sensing processing method and device for broadband signals based on deep learning

Technical Field

The present invention relates to the technical field of broadband signal processing, and in particular to a method, device, intelligent terminal and storage medium for compressed sensing processing of broadband signals based on deep learning.

Background Art

With the explosive growth of electromagnetic equipment and information systems, static spectrum management methods are no longer effective. A large number of frequency bands have low utilization rates, and there are few unallocated frequency bands suitable for data transmission. The contradiction between local tightness and overall idleness of wireless spectrum resources is becoming more and more obvious. At the same time, with the rapid development of mobile services and the advent of the Internet of Things era, the technical routes and key technologies adopted by future cellular networks require at least hundreds of megahertz of transmission bandwidth, hundreds of transmission antennas, and ultra-densely deployed base stations to support massive users.

Compressed Sensing (CS) theory provides a new solution for the acquisition and processing of broadband signals. In the prior art: the current mainstream compressed sampling frameworks include Multi-Coset Sampling (MCS), Random Demodulator Sampling (RDS), and Modulated Wideband Converter (MWC). Among them, the multi-coset sampling technology is a sub-Nyquist sampling technology with periodic non-uniform sampling, which can realize the compressed sampling of broadband signals through multiple low-speed ADCs with the same sampling rate but different sampling start times. The broadband spectrum sensing scheme based on multi-coset sampling in the prior art realizes sub-Nyquist sampling of broadband signals through the compressed sampling mode of the low-rate multi-channel architecture, and then restores the multi-band signal through the greedy algorithm to estimate the occupied channel position, thereby realizing the reconstruction of the input broadband signal. In this scheme, when the sampling rate and signal-to-noise ratio of the input broadband signal are low, the reconstruction accuracy of the signal still has a lot of room for improvement.

That is, the compressed sensing processing of broadband signals in the existing technology still has deficiencies in reconstruction accuracy, robustness, etc., and there are problems such as poor reconstruction performance for broadband spectrum sensing based on sub-Nyquist sampling.

Therefore, the prior art still needs to be improved and enhanced.

Summary of the invention

The technical problem to be solved by the present invention is that, in view of the above-mentioned defects of the prior art, a compressed sensing processing method, device, intelligent terminal and storage medium for broadband signals based on deep learning are provided. The present invention can solve the problems such as poor reconstruction performance of broadband spectrum sensing based on sub-Nyquist sampling.

In order to solve the above technical problems, the first aspect of the present invention provides a compressed sensing processing method for broadband signals based on deep learning, the method comprising:

Obtaining an input broadband signal, and performing multiple set sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence;

Preprocessing the obtained discrete sampling sequence and the input broadband signal and constructing a data set, wherein the data set includes: signal data of a training set and signal data of a test set;

Design a broadband signal reconstruction neural network model ADMM-net based on deep learning and initialize ADMM-net;

In the model training stage, the signal data of the training set is input into the ADMM-net, and the loss function is continuously minimized through the optimization algorithm to obtain the optimal neural network parameters;

During the testing phase or application phase, the signal data of the test set is input into the trained ADMM-net model to obtain the reconstructed broadband signal.

The compressed sensing processing method for broadband signals based on deep learning, wherein the step of obtaining the input broadband signal and performing multiple set sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence comprises:

采用多倍集采样在t＝(mL+c_i)T,i＝1,...,p,

时刻进行采样得到离散采样序列x_ci[n]，其中T表示输入的宽带信号的奈奎斯特采样时间周期间隔，与奈奎斯特采样定理相比，多倍集采样周期间隔为其L倍，故采样频率降低为奈奎斯特采样频率的1/L。Get the input broadband signal x(t) and determine the sampling mode through the sequence forward selection algorithm

Using multiple sampling at t = (mL + c _i )T, i = 1, ..., p,

Sampling is performed at every moment to obtain a discrete sampling sequence x _ci [n], where T represents the Nyquist sampling time period interval of the input broadband signal. Compared with the Nyquist sampling theorem, the multiple set sampling period interval is L times of it, so the sampling frequency is reduced to 1/L of the Nyquist sampling frequency.

The compressed sensing processing method of broadband signals based on deep learning, wherein the step of preprocessing the obtained discrete sampling sequence and the input broadband signal and constructing a data set comprises:

Perform discrete Fourier transform on the obtained discrete sampling sequence x _ci [n] to obtain the frequency domain representation of the sampling sequence

及对应的输入的宽带信号的频域信号X[k]组成数据集，其中训练集和测试集的数量为设定比例。At the same time, the frequency domain representation of the obtained sampling sequence is

And the frequency domain signal X[k] of the corresponding input broadband signal constitutes a data set, where the number of training sets and test sets is a set ratio.

The compressed sensing processing method for broadband signals based on deep learning, wherein the steps of designing a broadband signal reconstruction neural network model ADMM-net based on deep learning and initializing ADMM-net include:

Using the ADMM algorithm as a compressed sensing signal reconstruction algorithm, a deep learning-based scalable deep ADMM-net is designed to reconstruct broadband signals.

Initialize the parameters of the ADMM-net model to the corresponding parameter values in the ADMM algorithm.

The compressed sensing processing method for broadband signals based on deep learning, wherein, in the model training stage, the signal data of the training set is input into the ADMM-net, and the step of continuously minimizing the loss function through the optimization algorithm to obtain the optimal neural network parameters includes:

输入上述初始化的ADMM-net，得到初始重构的频域信号X_ir[k]；In the model training phase, the frequency domain representation of the sampling sequence of the training set is

Input the above initialized ADMM-net to obtain the initial reconstructed frequency domain signal _Xir [k];

Constructing a loss function E(θ) between the initial reconstructed frequency domain signal _Xir [k] and the frequency domain signal X[k] of the input broadband signal, where θ is a relevant neural network parameter;

The optimization algorithm is used to continuously minimize the loss function E(θ) to obtain the optimal neural network parameter θ, thereby obtaining the trained ADMM-net model.

The method for compressing broadband signals based on deep learning, wherein the step of inputting the signal data of the test set into the trained ADMM-net model to obtain the reconstructed broadband signal in the test phase or the application phase comprises:

输入所述训练后的ADMM-net模型进行处理，得到重构的频域信号X_r[k]。In the testing phase or application phase, the frequency domain representation of the sampling sequence of the test set is

The trained ADMM-net model is input for processing to obtain a reconstructed frequency domain signal X _r [k].

A compressed sensing processing device for broadband signals based on deep learning, wherein the device comprises:

The multiple sampling processing module is used to obtain the input broadband signal and perform multiple sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence;

A preprocessing module is used to preprocess the obtained discrete sampling sequence and the input broadband signal and construct a data set;

Set up and initialize the module, design a broadband signal reconstruction neural network model ADMM-net based on deep learning, and initialize ADMM-net;

A neural network training module is used to input the signal data of the training set into the ADMM-net during the model training phase, and continuously minimize the loss function through an optimization algorithm to obtain the optimal neural network parameters;

The test and application module is used to input the signal data of the test set into the trained ADMM-net model in the test phase or the application phase to obtain the reconstructed broadband signal.

An intelligent terminal includes a memory and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by one or more processors to include a method for executing one of the methods described.

A non-temporary computer-readable storage medium, characterized in that when the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute any one of the methods described.

Beneficial effect: Compared with the prior art, the present invention provides a method for compressing sensing of broadband signals based on deep learning. Based on the sparse spectrum characteristics of broadband signals, the present invention conducts research on compressing sensing of broadband signals based on deep learning on a simulated communication signal model. First, based on multiple set sampling, the input broadband signal is compressed and sampled in the time domain to obtain a discrete sampling sequence, and then the obtained discrete sampling sequence and the input broadband signal are preprocessed and a data set is constructed. At the same time, combined with deep learning, a scalable deep broadband signal reconstruction neural network model ADMM-net is designed to reconstruct the input broadband signal. In the model training stage, the signal data of the training set is input into the above-mentioned initialized ADMM-net to obtain the initial reconstructed frequency domain signal, and then the loss function between the initial reconstructed frequency domain signal and the frequency domain signal of the input broadband signal is constructed, and the optimization algorithm is used to continuously minimize the loss function, thereby obtaining the optimal neural network parameters. In the testing stage or the application stage, the signal data of the test set is input into the trained ADMM-net model to obtain the reconstructed broadband signal. Numerical analysis of experimental results shows that compared with commonly used reconstruction algorithms, ADMM-net combined with deep learning in the present invention has higher reconstruction accuracy under lower sampling rate and lower signal-to-noise ratio conditions after network model training, and has certain generalization ability.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a diagram of the compressed sensing theoretical framework.

FIG2 is a flowchart of a specific implementation of a method for compressive sensing of broadband signals based on deep learning provided in an embodiment of the present invention.

FIG3 is a multiple set sampling framework diagram of a compressed sensing processing method for broadband signals based on deep learning in a specific application embodiment of the present invention.

FIG4 is a schematic diagram of spectrum division and transfer of an input broadband signal of a compressed sensing processing method for broadband signals based on deep learning according to an application embodiment of the present invention.

FIG5 is a data flow diagram of ADMM-net of a compressed sensing processing method for broadband signals based on deep learning according to a specific application embodiment of the present invention.

6 is a diagram showing the reconstruction results of the three algorithms SOMP, ADMM and ADMM-net under different signal-to-noise ratio conditions, with the number of channels L=64 and the source channel k=5 in a compressed sensing processing method for broadband signals based on deep learning in a specific application embodiment of the present invention.

FIG7 is a principle block diagram of a compressed sensing processing device for broadband signals based on deep learning provided in an embodiment of the present invention.

FIG8 is a block diagram of the internal structure of the smart terminal provided in an embodiment of the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

With the explosive growth of electromagnetic equipment and information systems, static spectrum management methods are no longer effective. A large number of frequency bands have low utilization rates, and there are few unallocated frequency bands suitable for data transmission. The contradiction between local tightness and overall idleness of wireless spectrum resources is becoming more and more obvious. At the same time, with the rapid development of mobile services and the advent of the Internet of Things era, the technical routes and key technologies adopted by future cellular networks require at least hundreds of megahertz of transmission bandwidth, hundreds of transmission antennas, and ultra-densely deployed base stations to support massive users. According to the Nyquist sampling theorem, the minimum distortion-free sampling rate of the signal must be greater than or equal to twice the maximum bandwidth of the signal. With the increase of signal bandwidth, it becomes increasingly difficult to achieve high-speed sampling of broadband signals and storage, transmission and real-time processing of large-capacity data, which brings huge challenges to traditional communication systems.

Compressed Sensing (CS) theory provides a new solution for the acquisition and processing of broadband signals. Compared with the traditional Nyquist sampling system that first samples at a high rate and then compresses the data, compressed sensing directly compresses and samples sparse or compressible signals at a sub-Nyquist sampling rate, and then combines the reconstruction algorithm to reconstruct the original signal. Its theoretical framework is shown in Figure 1: original signal-sparse transformation-compressed sampling-signal reconstruction. Compressed sensing significantly reduces the requirements of hardware equipment such as storage, transmission, and analysis and processing for signal sampling in communication systems, thus bringing new opportunities to solve the challenges of high cost, low efficiency, and information redundancy that are difficult to overcome with traditional sampling methods.

In broadband spectrum sensing based on sub-Nyquist sampling, the selection of compressed sampling framework and the design of sparse signal reconstruction algorithm play a major role. Among them, the compressed sampling framework determines the way of compressed sampling of broadband signals, which is usually related to the design of random observation matrix, thus affecting the subsequent signal reconstruction process. The reconstruction algorithm has the greatest impact on the entire broadband spectrum sensing, which is related to the reconstruction complexity and accuracy of the signal and determines the performance of the entire broadband spectrum sensing. Therefore, it is very important to study and design a combination of a stable compressed sampling framework with low reconstruction complexity and a reconstruction algorithm to improve the performance of broadband spectrum sensing.

In the existing technology: the current mainstream compressed sampling frameworks include Multi-Coset Sampling (MCS), Random Demodulator Sampling (RDS), and Modulated Wideband Converter (MWC). Among them, the multi-coset sampling technology is a sub-Nyquist sampling technology with periodic non-uniform sampling. It can realize the compressed sampling of broadband signals through multiple low-speed ADCs with the same sampling rate but different sampling start times. Compared with other time-based ADC interleaved sampling schemes, multi-coset sampling only needs to select p<L channels out of L channels for sampling. Its sampling matrix has a low dimension, which reduces the sampling rate of the compressed sampling system. Its front-end circuit is simple and the cost of hardware implementation is lower, but its main difficulty lies in the precise design of sampling delay. The reconstruction algorithm is essentially to reconstruct the original signal by solving an underdetermined linear equation. Currently, the commonly used reconstruction algorithms include greedy algorithms, such as the Simultaneous Orthogonal Matching Pursuit (SOMP) algorithm, and convex relaxation algorithms, such as the Alternating Direction Method of Multipliers (ADMM) algorithm. Among them, the greedy algorithm has a lower computational complexity, but the reconstruction performance is poor when the observation value is small, and the amount of calculation gradually increases with the increase of the observation value. The convex relaxation algorithm can use a small number of samples to achieve approximate reconstruction of the input broadband signal, but its reconstruction complexity is relatively large. In the prior art, a broadband spectrum sensing scheme based on multiple set sampling is proposed. Through the compressed sampling mode of the low-rate multi-channel architecture, the sub-Nyquist sampling of the broadband signal is realized, and then the multi-band signal is restored by the greedy algorithm to estimate the occupied channel position, thereby realizing the reconstruction of the input broadband signal. In this scheme, when the sampling rate and signal-to-noise ratio of the input signal are low, the reconstruction accuracy of the signal still has a lot of room for improvement.

In order to solve the problems of the prior art, this embodiment provides a method for compressing and sensing broadband signals based on deep learning. The present invention is based on the sparse spectrum characteristics of broadband signals and conducts research on compressing and sensing broadband signals based on deep learning on a simulated communication signal model. First, based on multiple set sampling, the input broadband signal is compressed and sampled in the time domain to obtain a discrete sampling sequence, and then the obtained discrete sampling sequence and the input broadband signal are preprocessed and a data set is constructed. At the same time, combined with deep learning, a scalable deep broadband signal reconstruction neural network model ADMM-net is designed to reconstruct the input broadband signal. In the model training stage, the signal data of the training set is input into the above-mentioned initialized ADMM-net to obtain the initial reconstructed frequency domain signal, and then the loss function between the initial reconstructed frequency domain signal and the frequency domain signal of the input broadband signal is constructed, and the loss function is minimized by using an optimization algorithm to obtain the optimal neural network parameters. In the testing stage or the application stage, the signal data of the test set is input into the trained ADMM-net model to obtain a reconstructed broadband signal.

Numerical analysis of experimental results shows that compared with commonly used reconstruction algorithms, the broadband signal reconstruction neural network model ADMM-net combined with deep learning in the present invention has higher reconstruction accuracy under lower sampling rate and lower signal-to-noise ratio conditions after network model training, and has certain generalization ability.

Exemplary Methods

The method of this embodiment can be applied to a smart terminal. When it is specifically implemented, as shown in FIG2 , a method for compressing a broadband signal based on deep learning provided by an embodiment of the present invention specifically includes the following steps:

Step S100, obtaining an input broadband signal, and performing multiple set sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence;

When the present invention is implemented, the input broadband signal is first obtained, the sampling mode is determined by a sequence forward selection algorithm, and multiple set sampling is performed in the time domain to obtain a discrete sampling sequence.

Step S200, preprocessing the obtained discrete sampling sequence and the input broadband signal and constructing a data set, wherein the data set includes: signal data of a training set and signal data of a test set;

In this step, the obtained discrete sampling sequence is subjected to discrete Fourier transform (DFT transform) to obtain the frequency domain representation of the sampling sequence;

At the same time, the obtained frequency domain representation of the sampling sequence and the corresponding frequency domain signal of the input broadband signal are combined into a data set, wherein the number of training sets and test sets is a set ratio.

Step S300, designing a broadband signal reconstruction neural network model ADMM-net based on deep learning, and initializing ADMM-net;

In this step, the ADMM algorithm is used as the signal reconstruction algorithm for compressed sensing, and a scalable deep broadband signal reconstruction neural network model ADMM-net based on deep learning is designed to reconstruct broadband signals;

Initialize the parameters of the ADMM-net model to the corresponding parameter values in the ADMM algorithm.

Step S400: In the model training phase, the signal data of the training set is input into the ADMM-net, and the loss function is continuously minimized through the optimization algorithm to obtain the optimal neural network parameters;

In this step, during the model training phase, the frequency domain representation of the sampling sequence of the training set is input into the above-mentioned initialized ADMM-net to obtain the initial reconstructed frequency domain signal;

Constructing a loss function between the initially reconstructed frequency domain signal and the frequency domain signal of the input broadband signal;

The conjugate gradient method (CG) is used to continuously minimize the loss function to obtain the optimal neural network parameters, thereby obtaining the trained ADMM-net model.

Step S500: In the testing phase or the application phase, the signal data of the test set is input into the trained ADMM-net model to obtain a reconstructed broadband signal.

In this step, during the testing phase or the application phase, the frequency domain representation of the sampling sequence of the test set is input into the trained ADMM-net model for processing to obtain a reconstructed frequency domain signal.

The method of the present invention is further described in detail below by taking more detailed examples:

In the embodiment of the present invention, the step S100 of acquiring the input broadband signal and performing multiple set sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence is specifically implemented as follows:

时刻进行采样得到离散采样序列，其中T表示输入的宽带信号的奈奎斯特采样时间周期间隔，与奈奎斯特采样定理相比，多倍集采样周期间隔为其L倍，故其采样频率降低为奈奎斯特采样频率的1/L，图3为本发明实施例的多倍集采样框架图。集合

表示从{0,1,2,…,L-1}中选择的p<L个不同的整数，在多倍集采样中称为(L,p)采样模式。则第i个离散采样序列x_ci[n]可定义为公式(1)。Get the input broadband signal x(t), use multiple sampling at t＝(mL+ _ci )T,i＝1,2,…,p,

The discrete sampling sequence is obtained by sampling at the time, where T represents the Nyquist sampling time period interval of the input broadband signal. Compared with the Nyquist sampling theorem, the multiple set sampling period interval is L times of it, so its sampling frequency is reduced to 1/L of the Nyquist sampling frequency. FIG3 is a multiple set sampling framework diagram of an embodiment of the present invention.

represents p<L different integers selected from {0,1,2,…,L-1}, which is called (L,p) sampling mode in multiple set sampling. Then the i-th discrete sampling sequence x _ci [n] can be defined as formula (1).

i＝1,2,…,p，

表示第i个通道的采样时延，

表示奈奎斯特采样频率。in

i＝1,2,…,p，

represents the sampling delay of the i-th channel,

Represents the Nyquist sampling frequency.

In the embodiment of the present invention, a sequence forward selection algorithm is used to determine the sampling mode.

Furthermore, in the embodiment of the present invention, the specific implementation of step S200, preprocessing the obtained discrete sampling sequence and the input broadband signal and constructing a data set is as follows:

即In the present invention, as described in step S200, the sampled sequence x _ci [n] is subjected to discrete Fourier transform (DFT transform) to obtain the frequency domain representation of the i-th sampling sequence

Right now

Wherein k _c =0,1,2,…,M-1, i=1,2,…,p, and M represents the number of sampling points obtained by each sampling channel.

Assume that the input broadband signal x(t) is uniformly sampled at a time interval of T at the Nyquist sampling frequency _fs , and a discrete sequence x[n] with a total length of K is obtained. Then the sequence x[n] is subjected to a discrete Fourier transform to obtain the frequency domain signal X[k] of the input broadband signal, that is,

Where k = 0, 1, 2,…, K-1, K represents the number of points in the discrete sequence x[n].

i＝1,2,…,p及对应的输入的宽带信号的频域信号X[k]组成数据集，其中训练集和测试集的数量为设定比例(4:1)。训练集用于训练网络模型，测试集用于测试网络模型的重构性能及评估网络模型的泛化能力。The obtained sampling sequence is represented in the frequency domain

i＝1,2,…,p and the corresponding frequency domain signal X[k] of the input broadband signal constitute a data set, where the number of training sets and test sets is a set ratio (4:1). The training set is used to train the network model, and the test set is used to test the reconstruction performance of the network model and evaluate the generalization ability of the network model.

与输入的宽带信号的频域信号X[k]之间的关系，可得多倍集采样输入输出的矩阵表达式为Frequency domain representation of sampling sequence based on formulas (2) and (3)

The relationship between the frequency domain signal X[k] of the input broadband signal can be expressed as the matrix expression of the multiple set sampling input and output:

Y＝AX(4)

Where Y∈C ^p×M represents the frequency domain representation matrix of the sampling sequence, which is expressed as:

表示第i个通道采样序列的频域表示，i＝1,2,…,p，o表示Hadamard乘积；A∈C^p×L表示多倍集采样的测量矩阵，其组成元素的表达式为：

represents the frequency domain representation of the i-th channel sampling sequence, i = 1, 2, ..., p, o represents the Hadamard product; A∈C ^p×L represents the measurement matrix of multiple set sampling, and the expression of its constituent elements is:

T represents the Nyquist sampling time period interval of the input broadband signal; X∈C ^L×M represents the frequency domain representation matrix of the input broadband signal. The frequency domain signal X[k] of the input broadband signal is divided into L channels in units of 1/LT, and then the L-1 channels outside the first channel are moved to the left in units of 1/LT. The spectrum division transfer diagram is shown in Figure 4. The relationship expression between X and the frequency domain signal X[k] of the input broadband signal can be obtained, that is,

Where L represents the number of channels into which the spectrum is divided, M represents the number of signal points in each channel, and K=L·M represents the number of discrete signal points of the input broadband signal based on the Nyquist sampling frequency.

Furthermore, the present invention is about step S300, designing a broadband signal reconstruction neural network model ADMM-net based on deep learning, and initializing ADMM-net as follows:

Specifically, the present invention uses a convex relaxation method to reconstruct the input broadband signal, and its regularized unconstrained model can be expressed as:

Where X represents the frequency domain representation matrix of the input broadband signal to be reconstructed, A is the measurement matrix of the multiple set sampling, Y represents the frequency domain representation matrix of the sampling sequence, μ is the regularization parameter, and the auxiliary variable Z = [Z[1], Z[2], ..., Z[p]].

The augmented Lagrangian expansion L _P (X, Z, α) of equation (8) is:

Where α represents the Lagrange multiplier of the dual variable and ρ represents the Lagrange multiplier coefficient.

Using the ADMM algorithm as the signal reconstruction algorithm for compressed sensing, we can alternately optimize and solve {X, Z, α} through the three sub-problems separated by the above formula, and get:

分别表示重构信号操作、辅助变量更新操作及对偶变量更新操作的相关参数。Where k represents the kth iteration, X ^(k) represents the frequency domain representation matrix reconstructed at the kth iteration, Z ^(k) represents the auxiliary variable Z updated at the kth iteration, and M ^(k) represents the dual variables α, ρ, η and

They represent the relevant parameters of the reconstruction signal operation, the auxiliary variable update operation and the dual variable update operation respectively.

Combining the ideas of deep learning and data flow graphs, the iterative solutions of the three sub-problems of the above-mentioned ADMM algorithm can be designed into a data flow graph, thereby expanding into an alternating direction multiplier network that can be set to any number of layers, and its data flow graph is shown in Figure 5, which is a data flow graph of the ADMM-net model of an embodiment of the present invention. The data flow graph consists of nodes corresponding to different operations in the ADMM algorithm, and directed edges correspond to data flows between operations. In the kth stage of the figure, the three types of operations in the ADMM algorithm map three types of nodes, which are respectively defined as signal reconstruction layer (X ^(k) ), nonlinear transformation layer (Z ^(k) ) and multiplication update layer (M ^(k) ). The entire data flow graph is a multiple repetition of the above stages, corresponding to continuous iterations in the ADMM algorithm.

将其初始化为ADMM算法中相应的参数值。For the parameters of each layer in the ADMM-net model

Initialize it to the corresponding parameter value in the ADMM algorithm.

In a further embodiment, in step S400 of the present invention, in the model training stage, the signal data of the training set is input into the ADMM-net, and the loss function is continuously minimized by the optimization algorithm to obtain the optimal neural network parameters:

输入上述初始化的ADMM-net，得到初始重构的频域信号X_ir[k]。In the model training phase, the frequency domain representation of the sampling sequence of the training set is

Input the above initialized ADMM-net to obtain the initial reconstructed frequency domain signal _Xir [k].

In the embodiment of the present invention, the loss function of the initial reconstructed frequency domain signal _Xir [k] and the frequency domain signal X[k] of the input broadband signal for constructing the training set is defined as:

表示训练集N个信号数据中第i个信号数据的初始重构频域信号，(Y,θ)分别表示采样序列的频域表示矩阵Y和网络参数θ，其中θ包括信号重构层(X^(k))、非线性变换层(Z^(k))和乘法更新层(M^(k))的参数

X⁽ⁱ⁾表示训练集第i个信号数据的输入的宽带信号的频域信号。in

represents the initial reconstructed frequency domain signal of the i-th signal data in the training set N signal data, (Y, θ) represents the frequency domain representation matrix Y of the sampling sequence and the network parameter θ, respectively, where θ includes the parameters of the signal reconstruction layer (X ^(k) ), the nonlinear transformation layer (Z ^(k) ) and the multiplication update layer (M ^(k) )

X ⁽ⁱ⁾ represents the frequency domain signal of the input broadband signal of the i-th signal data in the training set.

In the model training of the embodiment of the present invention, the optimization algorithm of the network parameters is the conjugate gradient method (CG), and the gradient descent direction is determined by the Armijo criterion of the inexact one-dimensional search method. The conjugate gradient method is used to continuously minimize the loss function E(θ) to obtain the optimal neural network parameter θ, thereby obtaining the trained ADMM-net model.

Furthermore, in step S500 of the embodiment of the present invention, in the test phase or the application phase, the signal data of the test set is input into the trained ADMM-net model to obtain the reconstructed broadband signal, specifically:

输入所述训练后的ADMM-net模型进行处理，得到重构的频域信号X_r[k]。In the embodiment of the present invention, in the test phase or application phase, the trained ADMM-net model is used to test the reconstruction effect of broadband signals under different signal-to-noise ratio conditions.

The trained ADMM-net model is input for processing to obtain a reconstructed frequency domain signal X _r [k].

In the test phase of the present invention, the mean square error (MSE) is used to represent the final reconstruction error of the frequency domain signal:

Wherein, X _r [k] represents the frequency domain signal of the reconstructed wideband signal, and X[k] represents the frequency domain signal of the input wideband signal.

As can be seen from the above, the compressed sensing processing method of broadband signals based on deep learning in the embodiment of the present invention is mainly divided into five steps. The first is to obtain the input broadband signal, and perform multiple set sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence; the second is to preprocess the obtained discrete sampling sequence and the input broadband signal and construct a data set; the third is to design a broadband signal reconstruction neural network model ADMM-net based on deep learning, and initialize ADMM-net; the fourth is to input the signal data of the training set into the ADMM-net in the model training stage, and continuously minimize the loss function through the optimization algorithm to obtain the optimal neural network parameters; the fifth is to input the signal data of the test set into the trained ADMM-net model in the testing stage or the application stage to obtain the reconstructed broadband signal.

The implementation method of the present invention is: 300 input broadband signals are sampled multiple times at two different compression sampling rates of p/L=17% and p/L=42%, and then a data set is formed after preprocessing, wherein the number of data pairs of the training set and the test set is a set ratio (4:1), and the signal-to-noise ratio of the broadband signal of the training set is uniformly set in the range of [5,30] dB, and the signal-to-noise ratio of the broadband signal of the test set is set to 5 dB, 10 dB, 15 dB, 20 dB, 25 dB, 30 dB.

During the model training stage, the signal data of the training set obtained with two different compression sampling rates of p/L=17% and p/L=42% were respectively input into the ADMM-net model. The loss function was continuously minimized through the optimization algorithm to obtain the optimal neural network parameters, thereby obtaining the trained ADMM-net model.

In the testing phase or application phase, the signal data of the test set is input into the trained ADMM-net model to obtain the reconstruction results of the broadband signal under different signal-to-noise ratio conditions.

In the present invention, the channel L=64 of the broadband signal x(t) includes 5 source channels with different bandwidths. For signal data with two different compression sampling rates of p/L=17% and p/L=42% in the test set, under different signal-to-noise ratios, the signal reconstruction results of the three reconstruction algorithms SOMP, ADMM and ADMM-net are shown in Figure 6. Figure 6 is a schematic diagram of the reconstruction results of the three reconstruction algorithms SOMP, ADMM and ADMM-net under different signal-to-noise ratios when the number of channels L=64 and the source channel k=5 in an embodiment of the present invention.

It can be seen from the results of the above test set that compared with the commonly used reconstruction algorithms, the ADMM-net combined with deep learning in the present invention has higher reconstruction accuracy under lower sampling rate and lower signal-to-noise ratio conditions after network model training, and has a certain generalization ability.

And the present invention also has the following advantages:

1) Combining the idea of deep learning, for the first time, a scalable depth alternating direction multiplier network is used to reconstruct the frequency domain of broadband signals. The results show that the reconstruction performance of broadband signals can be improved.

2) By constructing a loss function between the network reconstruction result and the true value, the conjugate gradient method with the Armijo criterion as the gradient descent direction is used to continuously minimize the loss function, and the neural network parameters of the alternating direction multiplier network are trained, thereby improving the reconstruction performance of broadband signals and the generalization of the network model.

3) Using multiple set sampling and alternating direction multiplier network based on deep learning as the compressed sensing framework of broadband signals, the compressed sensing performance of broadband signals is improved under the conditions of lower sampling rate and lower signal-to-noise ratio.

Exemplary Devices

As shown in FIG. 7 , an embodiment of the present invention provides a compressed sensing processing device for broadband signals based on deep learning, the device comprising:

The multiple sampling processing module 410 is used to obtain the input broadband signal and perform multiple sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence;

A preprocessing module 420 is used to preprocess the obtained discrete sampling sequence and the input broadband signal and construct a data set;

A setting and initialization module 430 is used to design a broadband signal reconstruction neural network model ADMM-net based on deep learning and initialize ADMM-net;

A neural network training module 440 is used to input the signal data of the training set into the ADMM-net during the model training phase, and continuously minimize the loss function through an optimization algorithm to obtain the optimal neural network parameters;

The testing and application module 450 is used to input the signal data of the test set into the trained ADMM-net model during the testing phase or the application phase to obtain a reconstructed broadband signal, as described above.

Based on the above embodiments, the present invention also provides an intelligent terminal, whose principle block diagram can be shown in Figure 8. The intelligent terminal of the embodiment of the present invention can be a smart TV, and the intelligent terminal includes a processor, a memory, a network interface, and a display screen connected through a system bus. Among them, the processor of the intelligent terminal is used to provide computing and control capabilities. The memory of the intelligent terminal includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the intelligent terminal is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a compressed sensing processing method for broadband signals based on deep learning is implemented. The display screen of the intelligent terminal can be a liquid crystal display screen or an electronic ink display screen.

Those skilled in the art will understand that the principle block diagram shown in FIG8 is only a block diagram of a partial structure related to the solution of the present invention, and does not constitute a limitation on the smart terminal to which the solution of the present invention is applied. A specific smart terminal may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

In one embodiment, a smart terminal is provided, comprising a memory and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by one or more processors, wherein the one or more programs include instructions for performing the following operations:

Obtaining an input broadband signal, and performing multiple set sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence;

Preprocess the obtained discrete sampling sequence and the input broadband signal and construct a data set;

Design a broadband signal reconstruction neural network model ADMM-net based on deep learning and initialize ADMM-net;

In the model training stage, the signal data of the training set is input into the ADMM-net, and the loss function is continuously minimized through the optimization algorithm to obtain the optimal neural network parameters;

In the testing phase or the application phase, the signal data of the test set is input into the trained ADMM-net model to obtain a reconstructed broadband signal.

The step of obtaining an input broadband signal and performing multiple set sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence includes:

时刻进行采样得到离散采样序列x_ci[n]，其中T表示输入的宽带信号的奈奎斯特采样时间周期间隔，与奈奎斯特采样定理相比，多倍集采样周期间隔为其L倍，故采样频率降低为奈奎斯特采样频率的1/L。Get the input broadband signal x(t) and determine the sampling mode through the sequence forward selection algorithm

Sampling is performed at every moment to obtain a discrete sampling sequence x _ci [n], where T represents the Nyquist sampling time period interval of the input broadband signal. Compared with the Nyquist sampling theorem, the multiple set sampling period interval is L times of it, so the sampling frequency is reduced to 1/L of the Nyquist sampling frequency.

The step of preprocessing the obtained discrete sampling sequence and the input broadband signal and constructing a data set includes:

Perform discrete Fourier transform on the obtained discrete sampling sequence x _ci [n] to obtain the frequency domain representation of the sampling sequence

及对应的输入的宽带信号的频域信号X[k]组成数据集，其中训练集和测试集的数量为设定比例。At the same time, the frequency domain representation of the obtained sampling sequence is

And the frequency domain signal X[k] of the corresponding input broadband signal constitutes a data set, where the number of training sets and test sets is a set ratio.

The compressed sensing processing method for broadband signals based on deep learning, wherein the steps of designing a broadband signal reconstruction neural network model ADMM-net based on deep learning and initializing ADMM-net include:

Using the ADMM algorithm as a compressed sensing signal reconstruction algorithm, a deep learning-based scalable deep ADMM-net is designed to reconstruct broadband signals.

Initialize the parameters of the ADMM-net model to the corresponding parameter values in the ADMM algorithm.

Among them, in the model training stage, the signal data of the training set is input into the ADMM-net, and the loss function is continuously minimized by the optimization algorithm to obtain the optimal neural network parameters, including:

输入上述初始化的ADMM-net，得到初始重构的频域信号X_ir[k]；In the model training phase, the frequency domain representation of the sampling sequence of the training set is

Input the above initialized ADMM-net to obtain the initial reconstructed frequency domain signal _Xir [k];

Constructing a loss function E(θ) between the initial reconstructed frequency domain signal _Xir [k] and the frequency domain signal X[k] of the input broadband signal, where θ is a relevant neural network parameter;

The optimization algorithm is used to continuously minimize the loss function E(θ) to obtain the optimal neural network parameter θ, thereby obtaining the trained ADMM-net model.

Wherein, in the testing phase or the application phase, the step of inputting the signal data of the test set into the trained ADMM-net model to obtain the reconstructed broadband signal includes:

输入所述训练后的ADMM-net模型进行处理，得到重构的频域信号X_r[k]。In the testing phase or application phase, the frequency domain representation of the sampling sequence of the test set is

The trained ADMM-net model is input for processing to obtain a reconstructed frequency domain signal X _r [k].

Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and/or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

In summary, the present invention discloses a method, device, intelligent terminal and storage medium for compressed sensing of broadband signals based on deep learning, the method comprising: obtaining an input broadband signal, performing multiple set sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence; preprocessing the obtained discrete sampling sequence and the input broadband signal and constructing a data set; designing a broadband signal reconstruction neural network model ADMM-net based on deep learning, and initializing ADMM-net; in the model training stage, inputting the signal data of the training set into the ADMM-net, and continuously minimizing the loss function through the optimization algorithm to obtain the optimal neural network parameters; in the test stage or the application stage, inputting the signal data of the test set into the trained ADMM-net model to obtain a reconstructed broadband signal. The numerical analysis of the experimental results shows that compared with the commonly used reconstruction algorithm, the ADMM-net combined with deep learning in the present invention has a higher reconstruction accuracy under lower sampling rate and lower signal-to-noise ratio conditions after network model training, and has a certain generalization ability.

And the present invention also has the following advantages:

1) Combining the idea of deep learning, for the first time, a scalable depth alternating direction multiplier network is used to reconstruct the frequency domain of broadband signals. The results show that the reconstruction performance of broadband signals can be improved.

2) By constructing a loss function between the network reconstruction result and the true value, the conjugate gradient method with the Armijo criterion as the gradient descent direction is used to continuously minimize the loss function, and the neural network parameters of the alternating direction multiplier network are trained, thereby improving the reconstruction performance of broadband signals and the generalization of the network model.

3) Using multiple set sampling and alternating direction multiplier network based on deep learning as the compressed sensing framework of broadband signals, the compressed sensing performance of broadband signals is improved under the conditions of lower sampling rate and lower signal-to-noise ratio.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (7)

1. A method for compressed sensing processing of broadband signals based on deep learning, characterized in that the method comprises: Obtaining an input broadband signal, and performing multiple set sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence; Preprocessing the obtained discrete sampling sequence and the input broadband signal and constructing a data set, wherein the data set includes: signal data of a training set and signal data of a test set; Design a broadband signal reconstruction neural network model ADMM-net based on deep learning and initialize ADMM-net; In the model training stage, the signal data of the training set is input into the ADMM-net, and the loss function is continuously minimized through the optimization algorithm to obtain the optimal neural network parameters; In the testing phase or application phase, the signal data of the test set is input into the trained ADMM-net model to obtain the reconstructed broadband signal; The step of obtaining the input broadband signal and performing multiple set sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence also includes: Get the input broadband signal x(t) and determine the sampling mode through the sequence forward selection algorithm

Using multiple sampling

Sampling is performed at every moment to obtain a discrete sampling sequence x _ci [n], where T represents the Nyquist sampling time period interval of the input broadband signal. Compared with the Nyquist sampling theorem, the multiple set sampling period interval is L times of it, so the sampling frequency is reduced to 1/L of the Nyquist sampling frequency; Minimize the loss function using the conjugate gradient method using the Armijo criterion; In the model training stage, the signal data of the training set is input into the ADMM-net, and the loss function is continuously minimized by the optimization algorithm to obtain the optimal neural network parameters, including: In the model training phase, the frequency domain representation of the sampling sequence of the training set is

Input the above initialized ADMM-net to obtain the initial reconstructed frequency domain signal _Xir [k]; Constructing a loss function E(θ) between the initial reconstructed frequency domain signal _Xir [k] and the frequency domain signal X[k] of the input broadband signal, where θ is a relevant neural network parameter; The optimization algorithm is used to continuously minimize the loss function E(θ) to obtain the optimal neural network parameter θ, thereby obtaining the trained ADMM-net model; The loss function is defined as:

in

represents the initial reconstructed frequency domain signal of the i-th signal data in the training set N signal data, (Y, θ) respectively represent the frequency domain representation matrix Y of the sampling sequence and the network parameter θ, where X ⁽⁾ represents the frequency domain signal of the input broadband signal of the i-th signal data in the training set. 2. The method for compressive sensing of broadband signals based on deep learning according to claim 1, wherein the step of preprocessing the obtained discrete sampling sequence and the input broadband signal and constructing a data set comprises: Perform discrete Fourier transform on the obtained discrete sampling sequence x _ci [n] to obtain the frequency domain representation of the sampling sequence

At the same time, the frequency domain representation of the obtained sampling sequence is

and the frequency domain signal X[k] of the corresponding input broadband signal to form a data set, where the number of training sets and test sets is a set ratio; The input broadband signal x(t) is uniformly sampled at a time interval of T at the Nyquist sampling frequency _fs to obtain a discrete sequence x[n] with a total length of K. The sequence x[n] is then discretely Fourier transformed to obtain the frequency domain signal X[k] of the input broadband signal, that is,

Where k = 0, 1, 2,…, K-1, K represents the number of points in the discrete sequence x[n]. 3. The method for compressive sensing of broadband signals based on deep learning according to claim 1, wherein the steps of designing a signal reconstruction neural network model ADMM-net based on deep learning and initializing ADMM-net include: Using the ADMM algorithm as the signal reconstruction algorithm for compressed sensing, we designed a deep learning-based scalable signal reconstruction neural network model ADMM-net. Initialize the parameters of the ADMM-net model to the corresponding parameter values in the ADMM algorithm. 4. The method for compressive sensing of broadband signals based on deep learning according to claim 1, characterized in that, in the test phase or the application phase, the step of inputting the signal data of the test set into the trained ADMM-net model to obtain the reconstructed broadband signal comprises: In the testing phase or application phase, the frequency domain representation of the sampling sequence of the test set is

The trained ADMM-net model is input for processing to obtain a reconstructed frequency domain signal X _r [k]. 5. A compressed sensing processing device for broadband signals based on deep learning, characterized in that the device comprises: The multiple sampling processing module is used to obtain the input broadband signal and perform multiple sampling on the input broadband signal in the time domain to obtain a discrete sampling sequence; A preprocessing module, used to preprocess the obtained discrete sampling sequence and the input broadband signal and construct a data set, wherein the data set includes: signal data of a training set and signal data of a test set; The setup and initialization module is used to design the signal reconstruction neural network model ADMM-net based on deep learning and initialize ADMM-net; A neural network training module is used to input the signal data of the training set into the ADMM-net during the model training phase, and continuously minimize the loss function through an optimization algorithm to obtain the optimal neural network parameters; The test and application module is used to input the signal data of the test set into the trained ADMM-net model in the test phase or the application phase to obtain the reconstructed broadband signal; Get the input broadband signal x(t) and determine the sampling mode through the sequence forward selection algorithm

Using multiple sampling

Sampling is performed at every moment to obtain a discrete sampling sequence x _ci [n], where T represents the Nyquist sampling time period interval of the input broadband signal. Compared with the Nyquist sampling theorem, the multiple set sampling period interval is L times of it, so the sampling frequency is reduced to 1/L of the Nyquist sampling frequency; Minimize the loss function using the conjugate gradient method using the Armijo criterion; In the model training stage, the signal data of the training set is input into the ADMM-net, and the loss function is continuously minimized by the optimization algorithm to obtain the optimal neural network parameters, including: In the model training phase, the frequency domain representation of the sampling sequence of the training set is

Input the above initialized ADMM-net to obtain the initial reconstructed frequency domain signal _Xir [k]; Constructing a loss function E(θ) between the initial reconstructed frequency domain signal _Xir [k] and the frequency domain signal X[k] of the input broadband signal, where θ is a relevant neural network parameter; The optimization algorithm is used to continuously minimize the loss function E(θ) to obtain the optimal neural network parameter θ, thereby obtaining the trained ADMM-net model; The loss function is defined as:

in

represents the initial reconstructed frequency domain signal of the i-th signal data in the training set N signal data, (Y, θ) respectively represent the frequency domain representation matrix Y of the sampling sequence and the network parameter θ, where X ⁽⁾ represents the frequency domain signal of the input broadband signal of the i-th signal data in the training set. 6. An intelligent terminal, characterized in that it includes a memory and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by one or more processors, including for executing the method according to any one of claims 1 to 4. 7. A non-transitory computer-readable storage medium, characterized in that when the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the method as described in any one of claims 1 to 4.

Images