CN111970078A - Frame synchronization method for nonlinear distortion scene - Google Patents

Abstract

The invention discloses a frame synchronization method for a nonlinear distortion scene, comprising: collecting N _t M frames and N-long sample sequences y _i ⁽¹⁾ , y _i ⁽²⁾ , ... y _i ^(M) , i=1 ,2,...,N _t ; weighted superposition obtains the superimposed sample sequence y _i ^(S) , i=1,2,...,N _t ; preprocess the superimposed sample sequence _yi ^(S) to obtain the synchronization metric

i=1,2,...,N _t ; construct an ELM-based network, construct a label T _i according to the frame synchronization offset value of the received signal, i=1,2,...,N _t learn network parameters; use the learned ELM network The model learns to get an estimate of the frame sync offset

The invention can improve the frame synchronization performance in the nonlinear distortion scene, and compared with the traditional correlation method, the frame synchronization performance of the invention is greatly improved.

Description

A Frame Synchronization Method for Nonlinear Distortion Scenes

technical field

The invention relates to the technical field of wireless communication frame synchronization, in particular to a frame synchronization method for nonlinear distortion scenarios.

Background technique

As one of the important components in the communication system, the performance of the frame synchronization method directly affects the performance of the entire wireless communication system. However, there are inevitably nonlinear distortions in wireless communication systems, such as high-efficiency power amplifier distortion, analog-to-digital or digital-to-analog converter distortion, and distortion caused by unbalanced I/Q channels, and so on. In addition, next-generation wireless communication systems (such as 6G systems) need to use low-cost, low-resolution devices (such as power amplifiers, AD samplers) to avoid excessively expensive transceivers, resulting in particularly prominent nonlinear distortion. Traditional frame synchronization methods (such as correlation method) and newer frame synchronization methods have not considered nonlinear distortion scenarios, which makes them difficult to apply under nonlinear distortion conditions. Machine learning has excellent learning ability for nonlinear distortion. However, frame synchronization technology based on machine learning is rarely studied and has not achieved good synchronization performance, which needs to be improved urgently.

To this end, the present invention utilizes a machine learning method and develops inter-frame correlation prior information to form a frame synchronization method that improves the performance of frame synchronization error probability. At the receiving end, the frame is first preprocessed by weighted superposition, the prior information of inter-frame correlation is developed, and frame synchronization metric features are initially captured; then, an ELM frame synchronization network is constructed to perform offline training for the estimation of frame synchronization offset; finally, The frame synchronization offset is estimated online by combining the preprocessing and the learned ELM network parameters. For wireless communication scenarios with nonlinear distortion, such as 5G and 6G systems, the present invention can improve the error probability performance of its frame synchronization and promote the intelligent processing level of its frame synchronization, bringing many implementable solutions for the research of intelligent frame synchronization ,has great significance.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an intelligent frame synchronization method for nonlinear distortion scenes. Compared with the traditional correlation synchronization method, this method combines multi-frame weighted superposition and ELM network to effectively improve the frame synchronization performance under the nonlinear distortion system. .

The specific invention scheme is:

A frame synchronization method for a nonlinear distortion scene, comprising the following steps:

(a) collect N _t M frames of N-long sample sequences y _i ⁽¹⁾ , y _i ⁽²⁾ ,...y _i ^(M) , i=1, 2,...,N _t ;

(b ₎ Perform weighted stacking on y _i ^{( 1) ,} y _i ⁽²⁾ _{, ...}

(c) Preprocess the superimposed sample sequence _yi ^(S) to obtain a standard metric vector

(d) Constructing an ELM network, constructing labels T _i according to the frame synchronization offset value of the received signal, i=1, 2,...,N _t to learn network parameters;

(e) Using the learned ELM network model to learn the estimated value of frame synchronization offset

Further, in the step (a), the obtained M frame N-long sample sequence can be expressed as:

Among them, M and N are set according to engineering experience.

Further, the weighted superposition of the method step (b) is expressed as:

y _i ^(S) = μ ₁ y _i ⁽¹⁾ + μ ₂ y _i ⁽²⁾ +...+μ _M y _i ^(M) ;

The μ _i , i=1, 2, . . . , M are weighting coefficients; they are set according to the received signal-to-noise ratio of each frame.

Further, the preprocessing step described in the method step (c) is:

与长度为N_s的训练序列

作“互相关运算”后，得到“互相关度量”Γ_t，即：(c1) An observation sequence with an observation length of N _s in a training superimposed sample sequence y ^(S)

with a training sequence of length N _s

After "cross-correlation operation", the "cross-correlation measure" Γ _t is obtained, namely:

The observation length is N _s , which is set according to engineering experience;

表示y^(S)中t到t+N_s-1个元素；The t represents the starting index position of the observation stacking sequence, t∈[1,K], for example, t=1 means that the _Ns -long sample sequence is observed from the first element of the stacking sample sequence y ^(S) ;

Represents t to t+N _s -1 elements in y ^(S) ;

The K=NN _s +1 represents the size of the search window;

构成度量矢量

对N_t个度量矢量γ_i进行归一化处理，得到标准度量矢量

即：(c2) is measured by K correlations

composition metric vector

Normalize the N _t metric vectors γ _i to obtain a standard metric vector

which is:

The N _t is set according to engineering experience, and the ||γ _i || represents the Frobenius norm of the metric vector γ _i .

Further, the network model and parameters of the step (d) are:

输出层节点数为K，隐藏层采用sigmoid作为激活函数，将预处理后的标准度量矢量集合

作为输入；The ELM network model includes 1 input layer, 1 hidden layer, and 1 output layer. The number of nodes in the input layer is K, and the number of nodes in the hidden layer is

The number of nodes in the output layer is K, the hidden layer uses sigmoid as the activation function, and the preprocessed standard metric vector set is

as input;

The m is set according to engineering experience.

Further, the step of constructing the label described in the step (d) is:

According to the synchronization offset value τ _i , i=1,2,...,N _t forms a tag set

The labels T _i , i=1, 2,...,N _t are obtained by one-hot encoding according to the synchronization offset value τ _i , that is,

The τ _i is determined by the received signal _yi , and is collected according to a statistical channel model, or according to an actual scenario in combination with existing methods or devices.

Further, the offline training process of the step (d) specifically includes the following steps:

和偏置

依次将标准度量矢量

输入ELM网络，隐藏层输出

表示为：(d1) Generate weights according to random distribution

and bias

The standard metric vector in turn

Input to ELM network, hidden layer output

Expressed as:

The σ( ) represents the activation function sigmoid;

得到的N_t个隐藏层输出H_i构成隐藏层输出矩阵

根据隐藏层输出矩阵H和步骤(a3)中构建的标签集合T，求得输出权重

(d2) consists of N _t standard metric vectors

The obtained N _t hidden layer outputs H _i form the hidden layer output matrix

According to the hidden layer output matrix H and the label set T constructed in step (a3), the output weight is obtained

表示H的Moore–Penrose伪逆；said

represents the Moore–Penrose pseudoinverse of H;

(d3) Save the model parameters W, b and β.

Further, the online running process of the step (e) includes the following steps:

将

送入ELM网络模型中学习出输出向量

表示为：(e1) Receive M frames of N-long online sample sequences y _online ⁽¹⁾ , y _online ⁽²⁾ ,..., y _online ^(M) , and perform superposition preprocessing according to steps (b) and (c) to obtain an online standard metric vector

Will

Feed into the ELM network model to learn the output vector

Expressed as:

(e2) Find the index position of the maximum value of the square of the amplitude in the output vector O, that is, the estimated value of frame synchronization

The beneficial effect of the invention is that the frame synchronization performance under the nonlinear distortion system is improved.

Description of drawings

Fig. 1 is the schematic flow chart of the present invention;

Figure 2 is a flowchart of offline training of ELM network;

Figure 3 is a diagram of the online running process of the ELM network.

Detailed ways

The technical solutions of the present invention are further described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the following.

As shown in Figure 1, a frame synchronization method for a nonlinear distortion scene includes the following steps:

(a) Collect N _t M frames of N-long sample sequences y _i ⁽¹⁾ , y _i ⁽²⁾ ,...y _i ^(M) , i=1, 2,...,N _t ;

Specifically, in the method step (a), the obtained M frame N-length sample sequence can be expressed as:

Among them, M and N are set according to engineering experience.

(b ₎ Perform weighted stacking on y _i ^{( 1) ,} y _i ⁽²⁾ _{, ...}

Specifically, the weighted superposition of the method step (b) can be expressed as:

y _i ^(S) = μ ₁ y _i ⁽¹⁾ + μ ₂ y _i ⁽²⁾ +...+μ _M y _i ^(M) ;

The μ _i , i=1, 2, . . . , M are weighting coefficients; they are set according to the received signal-to-noise ratio of each frame.

Example 1: The weighting coefficients are set as follows:

Assuming M=3, the signal-to-noise ratios of the three frame signals are α ₁ , α ₂ , α ₃ respectively,

(c) Preprocess the superimposed sample sequence y _i ^(S) to obtain a synchronization metric

Specifically, the preprocessing step described in the method step (c) is:

与长度为N_s的训练序列

作“互相关运算”后，得到“互相关度量”Γ_t，即：(c1) An observation sequence with an observation length of N _s in a training superimposed sample sequence y ^(S)

with a training sequence of length N _s

After "cross-correlation operation", the "cross-correlation measure" Γ _t is obtained, namely:

The observation length is N _s , which is set according to engineering experience;

表示y^(S)中t到t+N_s-1个元素；The t represents the starting index position of the observation stacking sequence, t∈[1,K], for example, t=1 means that the _Ns -long sample sequence is observed from the first element of the stacking sample sequence y ^(S) ;

Represents t to t+N _s -1 elements in y ^(S) ;

The K=NN _s +1 represents the size of the search window;

构成度量矢量

对N_t个度量矢量γ_i进行归一化处理，得到标准度量矢量

即：(c2) is measured by K correlations

composition metric vector

Normalize the N _t metric vectors γ _i to obtain a standard metric vector

which is:

The N _t is set according to engineering experience, and the ||γ _i || represents the Frobenius norm of the metric vector γ _i .

(d) Constructing an ELM network, constructing labels T _i according to the frame synchronization offset value of the received signal, i=1, 2,...,N _t to learn network parameters;

In the embodiment of the present application, the network model and parameters of the method step (d) are:

输出层节点数为K，隐藏层采用sigmoid作为激活函数，将预处理后的标准度量矢量集合

作为输入；The ELM network model includes 1 input layer, 1 hidden layer, and 1 output layer. The number of nodes in the input layer is K, and the number of nodes in the hidden layer is

The number of nodes in the output layer is K, the hidden layer uses sigmoid as the activation function, and the preprocessed standard metric vector set is

as input;

The m is set according to engineering experience.

Specifically, the step of constructing the label described in the method step (d) is:

According to the synchronization offset value τ _i , i=1,2,...,N _t forms a tag set

The labels T _i , i=1, 2,...,N _t are obtained by one-hot encoding according to the synchronization offset value τ _i , that is,

The τ _i is determined by the received signal _yi , and is collected according to a statistical channel model, or according to an actual scenario in combination with existing methods or devices.

Example 2: An example of the label in said step (d) is as follows:

Assuming N=64, τ _i =5, N _t =10 ⁵ ,

training labels:

As shown in FIG. 2, in the embodiment of the present application, the offline training process of the method step (d) specifically includes the following steps:

和偏置

依次将标准度量矢量

输入ELM网络，隐藏层输出

表示为：(d1) Generate weights according to random distribution

and bias

The standard metric vector in turn

Input to ELM network, hidden layer output

Expressed as:

The σ( ) represents the activation function sigmoid;

得到的N_t个隐藏层输出H_i构成隐藏层输出矩阵

根据隐藏层输出矩阵H和步骤(d)中构建的标签集合T，求得输出权重

(d2) consists of N _t standard metric vectors

The obtained N _t hidden layer outputs H _i form the hidden layer output matrix

According to the hidden layer output matrix H and the label set T constructed in step (d), the output weight is obtained

表示H的Moore–Penrose伪逆；said

represents the Moore–Penrose pseudoinverse of H;

(d3) Save the model parameters W, b and β.

(e) Using the learned ELM network model to learn the estimated value of frame synchronization offset

As shown in Figure 3, in the embodiment of the present application, specifically, the online operation process of step (e) includes the following steps:

将

送入ELM网络模型中学习出输出向量

表示为：(e1) Receive M frames of N-long online sample sequences y _online ⁽¹⁾ , y _online ⁽²⁾ ,..., y _online ^(M) , and perform superposition preprocessing according to steps (b) and (c) to obtain an online standard metric vector

Will

Feed into the ELM network model to learn the output vector

Expressed as:

(e2) Find the index position of the maximum value of the square of the amplitude in the output vector O, that is, the estimated value of frame synchronization

It should be noted that those of ordinary skill in the art will realize that the embodiments described herein are intended to help readers understand the implementation method of the present invention, and it should be understood that the protection scope of the present invention is not limited to such special statements and examples. Those skilled in the art can make various other specific modifications and combinations without departing from the essence of the present invention according to the technical teachings disclosed in the present invention, and these modifications and combinations still fall within the protection scope of the present invention.

Claims (8)

1. A frame synchronization method for a nonlinear distortion scene is characterized by comprising the following steps:

(a) collecting N_tM frames of N long sample sequences y_i ⁽¹⁾,y_i ⁽²⁾,…y_i ^(M),i＝1,2,…,N_t；

(b) For y_i ⁽¹⁾,y_i ⁽²⁾,…y_i ^(M)Carrying out weighted superposition to obtain a superposed sample sequence y_i ^(S),i＝1,2,…,N_t

(c) For the superimposed sample sequence y_i ^(S)Preprocessing to obtain standard measurement vector

(d) Constructing an ELM network, and constructing a tag T according to the frame synchronization deviation value of the received signal_i,i＝1,2,…,N_tLearning network parameters;

(e) learning to obtain frame synchronization offset estimation value by using learning-obtained ELM network model

2. The method for frame synchronization of a non-linear distortion scene according to claim 1, wherein the sequence of samples of M frames N length in step (a) is represented as:

wherein M and N are set according to engineering experience.

3. The method for frame synchronization of a non-linearly distorted scene as claimed in claim 1, wherein the weighted overlap-add of step (b) is represented by:

y_i ^(S)＝μ₁y_i ⁽¹⁾+μ₂y_i ⁽²⁾+…+μ_My_i ^(M)；

the mu_iWhere i is 1,2, …, and M is a weighting coefficient, and is set according to the received snr of each frame.

4. The method for frame synchronization of a non-linear distortion scene as claimed in claim 1, wherein the preprocessing step of step (c) is:

(c1) one-time training superposition sample sequence y^(S)Middle observation length of N_sObservation sequence of

And length N_sTraining sequence of

After the cross-correlation operation, the cross-correlation measurement is obtained "_tNamely:

the observation length is N_sSetting according to engineering experience;

the t represents the initial index position of the observed superposition sequence, and t belongs to [1, K ]]For example, t-1 denotes a sequence y of samples from the superposition^(S)Begins to observe N_sA long sample sequence;

denotes y^(S)Middle t to t + N_s-1 element;

the K is N-N_s+1, representing the size of the search window;

(c2) measured by K correlations

Constructing a metric vector

To N_tIndividual measurement vector gamma_iNormalization processing is carried out to obtain a standard measurement vector

Namely:

said N is_tAccording to the engineering experience setting, the | | | gamma_i| represents the measurement vector γ_iFrobenius norm of (1).

5. The improvement method of claim 1, wherein said network model and parameters of step (d) are:

the ELM network model comprises 1 input layer, 1 hidden layer and 1 output layer, wherein the number of nodes of the input layer is K, and the number of nodes of the hidden layer is K

The number of nodes of an output layer is K, a hidden layer adopts sigmoid as an activation function, and preprocessed standard measurement vectors are integrated

As an input;

and m is set according to engineering experience.

6. The method for frame synchronization of a non-linear distortion scene as claimed in claim 1, wherein the step of constructing the label in step (d) is:

according to the synchronization deviation value tau_i,i＝1,2,…,N_tForming a set of tags

The label T_i,i＝1,2,…,N_tAccording to the synchronization deviation value tau_iObtained by one-hot coding, i.e.

Said tau_iFrom the received signal y_iAnd determining, and collecting according to a statistical channel model or according to an actual scene by combining the existing method or equipment.

7. The improvement method of claim 1, wherein the offline training process (d) comprises the following steps:

(d1) generating weights from random distributions

And bias

Sequentially combining the standard metric vectors

Input to ELM network, hidden layer output

Expressed as:

the σ (-) represents an activation function sigmoid;

(d2) from N_tIndividual metric vector

Obtained N_tA hidden layer output H_iConstructing hidden layer output matrices

Obtaining output weight according to hidden layer output matrix H and label set T constructed in step (a3)

The above-mentioned

Moore-Penrose pseudoinverse representing H;

(d3) model parameters W, b and β are saved.

8. The improvement method of claim 1, wherein the on-line operation process of step (e) comprises the steps of:

receiving M frames of N long online sample sequences y_online ⁽¹⁾,y_online ⁽²⁾,…,y_online ^(M)Performing superposition preprocessing according to the steps (b) and (c) to obtain an online standard measurement vector

Will be provided with

The vector is sent to an ELM network model to learn an output vector

Expressed as:

(e1) finding the index position of the maximum of the squared amplitudes in the output vector O, i.e. the frame synchronization estimate

Images