CN110830404A - Digital mobile forward signal quantization method based on vector linear prediction - Google Patents

Abstract

The invention discloses a digital mobile fronthaul signal quantization method based on vector linear prediction. The OFDM modulated signal is subjected to IFFT to obtain real part and imaginary part I/Q sampling signals, which are then constructed into a multi-dimensional vector set; The vector set is subjected to differential vector quantization, and a column quantized signal composed of codeword-index mapping sequence is output; the quantized signal is PAM-4 encoded, and then converted into an optical signal through electro-optical modulation; the optical signal is input into a single-mode fiber and transmitted to the radio frequency Extend the unit, recover the quantized signal after photoelectric detection and PAM-4 decoding; perform index-code word mapping and differential demodulation on the recovered quantized signal to recover the original differential vector quantized signal; quantize the recovered differential vector The signal is demodulated by the vector set, and the I/Q two-way sampled signal of the OFDM‑IFFT is recovered. On the premise of ensuring good performance of digital mobile fronthaul, the method of the invention greatly improves the spectral efficiency of transmission signals and enriches the realization method of digital mobile fronthaul.

Description

A Digital Mobile Fronthaul Signal Quantization Method Based on Vector Linear Prediction

technical field

The invention relates to the technical field of 5G mobile communication, in particular to a digital mobile fronthaul signal quantization method based on vector linear prediction.

Background technique

With the commercialization of the fifth-generation mobile communication (5G), broadband and high-speed 5G data transmission will provide strong support for some emerging industries such as the Internet of Things and VR/AR. Cloud Radio Access Network (C-RAN) is a centralized signal processing network architecture suitable for 5G. By reducing the number of base station computer rooms and installing low-cost simplified base stations on a large scale, the transmission distance of millimeter-wave signals with short transmission distances is reduced. full coverage. Digital mobile fronthaul can be regarded as a signal transmission method in the mobile fronthaul part of the C-RAN access network architecture. The base station room is a baseband unit pool that implements centralized digital processing of radio frequency signals. Simplified base station is actually the recovery and transmission of radio frequency signals. For remote radio units with other tasks, the baseband unit pool and the remote radio units are connected by single-mode optical fibers that transmit digital baseband signals. How to improve the spectral efficiency of digital baseband signals transmitted in optical fibers has always been one of the difficult problems that digital mobile fronthaul networks need to overcome.

As far as the current research situation is concerned, there are many kinds of signal quantization methods for realizing digital mobile fronthaul. The traditional signal quantization is mainly based on scalar quantization. As the most primitive analog signal digitization method, PCM adopts uniform or compressed scalar quantization to realize the process is relatively simple. DPCM performs differential processing on the signal before quantization, which greatly optimizes the quantization performance of PCM. In addition, the method of vector quantization has also begun to be applied to digital mobile fronthaul. Compared with scalar quantization, the traditional vector quantization technology based on k-means clustering has better quantization performance.

It should be pointed out that the number of quantization bits required for PCM quantization is large, which leads to a low spectral utilization rate of digital signals; although the use of DPCM technology for signal quantization improves the spectral utilization rate of PCM technology, it still needs to The Q two-way signal components are transmitted by time division multiplexing; while the vector quantization technology based on k-means clustering greatly reduces the quantization performance when the dynamic range of the signal is large. The invention provides a digital mobile fronthaul signal quantization method based on vector linear prediction.

SUMMARY OF THE INVENTION

In view of the above problems, the purpose of the present invention is to provide a digital mobile fronthaul based on vector linear prediction, which can greatly improve the spectral efficiency of the transmitted signal and enrich the implementation method of the digital mobile fronthaul under the premise of ensuring the good performance of the digital mobile fronthaul. Signal quantization method. The technical solution is as follows:

A digital mobile fronthaul signal quantization method based on vector linear prediction, comprising the following steps:

Step 1: perform IFFT on the OFDM modulated signal to obtain real part and imaginary part I/Q two-way sampling signals, and then construct the I/Q two-way sampling signals into a multi-dimensional vector set;

Step 2: use the differential vector quantization method based on vector linear prediction to perform differential vector quantization on the input multi-dimensional vector set, and output a column quantized signal composed of a codeword-index mapping sequence;

Step 3: The quantized signal is encoded by PAM-4, and then converted into an optical signal through electro-optical modulation;

Step 4: Input the optical signal into the single-mode fiber and transmit it to the remote radio unit, and then recover the quantized signal after photoelectric detection and PAM-4 decoding;

Step 5: perform index-codeword mapping and differential demodulation on the recovered quantized signal according to the inverse process of step 2 to recover the original differential vector quantized signal;

Step 6: demodulate the vector set on the recovered differential vector quantized signal according to the inverse process of step 1, and recover the I/Q two-channel sampling signal of the OFDM-IFFT.

Further, the specific process of performing differential vector quantization on the input multi-dimensional vector set using the differential vector quantization method based on vector linear prediction is:

Step 21: Set the parameters of vector linear prediction, including the length of the training vector sequence and the prediction order;

Step 22: offline training process: intercepting a multi-dimensional vector set of a specific length as a training sample sequence, and using the vector linear prediction method to obtain the required prediction coefficient matrix and optimal codebook;

Step 23: Online quantization process: using the obtained prediction coefficient matrix and the optimal codebook, the multi-dimensional vector set signal is input into the differential vector quantizer based on vector linear prediction, and the output quantized signal is each of the optimal codebooks. The index number sequence corresponding to the codeword, wherein the vector quantization adopts an algorithm based on k-means clustering.

Further, the specific process of the step 22 is:

Step 1) intercepting a multi-dimensional vector set with a certain length of N from the constructed multi-dimensional vector set as an offline training sequence X={s ₁ ,s ₂ ,...,s _N };

Step 2) Define a finite memory vector linear predictor with memory length p by predicting coefficient matrix set A={A ₁ ,A ₂ ,...,A _p }; A _j is a D×D matrix, j∈[1,p ], p is the order of the vector linear predictor; D is the dimension of the multi-dimensional vector set;

Step 3) For online quantization, the input of the vector quantizer is the predicted vector error expressed as:

是预测器的输出矢量值；where _sn is the input multidimensional vector,

is the output vector value of the predictor;

The p-order vector linear predictor obtains the predicted value of the input multi-dimensional vector s _n by the following formula through the sampled observations s _n-1 , s _n-2 ,...,s _np at the previous p times

选择使p阶预测器性能测度J最小的预测系数矩阵集A：Step 4) Define the performance measure J of the p-order predictor as:

Choose the prediction coefficient matrix set A that minimizes the performance measure J of the p-order predictor:

Define the D×D correlation matrix as

分别为D×1,1×D阶矩阵。Here i,j∈[1,p],

They are D×1 and 1×D order matrices, respectively.

According to the orthogonality rule, when the error components generated by the linear predictor are orthogonal to the observed values, we have

Substituting equations (1) and (2) into equation (3), we get

In the formula, v∈[1,p]; A _v is the corresponding vth sub-matrix in the prediction coefficient matrix set in step 2.

Formula (4) can be written as:

Formula (5) is represented by the following matrix equation:

For formula (6), the Levinson-Durbin algorithm is used to solve the coefficient matrix set A of the optimal linear predictor of the p-order stationary D-dimensional vector sequence in general;

Step 5) Use k-means clustering method to cluster the multi-dimensional vector set with the length of N to obtain the optimal codebook C; the length of the codebook is determined by the number of quantized bits, and the number of quantized bits Qb=log ₂ k , k is the number of clusters of k-means.

The beneficial effects of the present invention are: the present invention uses vector quantization instead of scalar quantization, which greatly improves the spectral efficiency of the digital mobile fronthaul link; uses vector linear prediction technology to perform differential quantization processing on vector signals to reduce the dynamic range of signals , to reduce the noise of the vector quantization of the signal; by changing the dimension of the vector signal, the order p of the vector predictor and other parameters, the signal quantization of the digital mobile fronthaul becomes more flexible.

Description of drawings

FIG. 1 is a block diagram of the implementation of a digital mobile fronthaul signal quantization method based on vector linear prediction of the present invention.

Figure 2 is a block diagram of VLP-VQ.

FIG. 3 is a schematic diagram of a digital mobile fronthaul signal quantization based on vector linear prediction.

FIG. 4 is a schematic diagram of quantization of two-dimensional vector signals: (a) VLP-OFDM sampling; (b) OFDM sampling.

FIG. 5 is a performance test diagram of a digital mobile fronthaul receiver based on vector linear prediction.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

As shown in FIG. 1, the present invention provides a digital mobile fronthaul signal quantization method based on vector linear prediction, comprising the following steps:

Step 1: perform IFFT on the OFDM modulated signal to obtain real part and imaginary part I/Q sampling signals, and then construct the I/Q two sampling signals into a multi-dimensional vector set.

L为采样信号长度。信号的采样率f_s可以表示为：f_s＝P_IFFT×ΔP，P_IFFT为IFFT/FFT点数，ΔP为OFDM的子载波间隔。The real part and imaginary part of the OFDM modulated signal after IFFT are respectively the sampled signals (100) of the I/Q channels. The sampled signals can be expressed as: S={s _O1 ,s _O2 ,...,s _OL }, where

L is the length of the sampled signal. The sampling rate f _s of the signal can be expressed as: f _s =P_IFFT×ΔP, where P_IFFT is the number of IFFT/FFT points, and ΔP is the subcarrier spacing of OFDM.

The sampling signal S constructs a multi-dimensional vector set 200 according to a certain rule according to the dimension D of the vector. For example, to construct a vector set with a dimension of 4, the specific vector set construction method is shown in formula (1):

Step 2: Use the differential vector quantization method based on vector linear prediction to perform differential vector quantization on the input multi-dimensional vector set, and output a column quantized signal composed of a codeword-index mapping sequence.

FIG. 2 is a block diagram showing the steps of vector signal quantization (VLP-VQ) 300 based on vector linear prediction for digital mobile fronthaul. First, intercept a multi-dimensional vector set 301 of a certain length in formula (1) for offline training to obtain the required prediction coefficient matrix and optimal codebook (302), and then use the obtained prediction coefficient matrix and optimal codebook 302 Perform online differential quantization on the multi-dimensional vector set, and obtain a differential vector quantized signal 303, the quantization value is determined by the code word in the optimal codebook, and finally the obtained differential vector quantized signal is subjected to code word-index number mapping to obtain a set of Quantized signal 304 in the range 1--k. Among them, k is the number of clusters of k-means, quantization bit Qb=log ₂ k, here, quantization bit/sample is used instead of quantization bit number, quantization bit/sample Qb/Sa=Qb/D. The implementation process of VLP-VQ is shown in the block diagram of Figure 3, and the specific reasoning process is as follows:

[1] For online quantization, the input to the vector quantizer is the predicted vector error, which can be expressed as:

是预测器的输出矢量值。where _sn is the input multidimensional vector,

is the output vector value of the predictor.

[2] Then, the predicted vector error e _n is quantized by a vector quantizer into The quantization error of the vector quantizer can be expressed as:

[3] At the receiving end, the reconstructed multidimensional vector set signal can be expressed as:

[4] Substitute equations (2) and (3) into equation (4), we can get

的差别，仅是预测矢量误差e_n的量化误差q_n。显然，这种预测矢量量化比直接对多维矢量进行矢量量化的效果要好。According to formula (5), the input multi-dimensional vector _sn and the reconstructed multi-dimensional vector

The difference is only the quantization error q _{n of the predicted vector error e n .} Obviously, this predictive vector quantization is better than direct vector quantization of multi-dimensional vectors.

[5] Before performing the above prediction vector quantization, we need to obtain the required prediction coefficient matrix set A and optimal codebook C according to the vector linear prediction method, where A={A ₁ ,A ₂ ,...,A _p }, p is the order of the vector linear predictor, and the optimal codebook C is obtained by clustering the input multi-dimensional vector set according to k-means clustering. A multi-dimensional vector set with a certain length of N is intercepted as an offline training sequence X={s ₁ , s ₂ ,...,s _N }, and the required A and C are obtained. For vector linear prediction, a finite-memory vector linear predictor with memory length p can be defined by a set of prediction coefficient matrices A, this p-order predictor is passed through the previous p observations s _n-1 , s _n-2 , ... , s _np , the predicted value of s _n is obtained from equation (6)

In formula (6), A _j is a D×D matrix.

对于给定记忆长度为p,如果矢量线性预测器所选择的系数矩阵集A能使J的测度最小，则称该矢量线性预测器是最优的。[6] If the squared error measure is used, the performance measure J of the p-order predictor can be defined as

For a given memory length p, the vector linear predictor is said to be optimal if the coefficient matrix set A selected by the vector linear predictor can minimize the measure of J.

[7] In order to describe the statistical properties of random vector sequences, it is necessary to determine the second-order statistical properties of the vector process {s _n }. The D×D correlation matrix is defined as follows:

easy to prove

的线性预测器，当且仅当它所产生的各误差分量与各观测值正交，它才是最佳的，所以：[8] According to the orthogonality rule, for the form such as

The linear predictor of is optimal if and only if the error components it produces are orthogonal to the observations, so:

[9] Substitute equations (2) and (6) into equation (8), we can get:

Equation (9) can also be written as:

[10] Equation (10) can be represented by the following matrix equation:

For formula (11), the Levinson-Durbin algorithm can be used to solve the coefficient matrix set A of the optimal linear predictor of the p-order stationary D-dimensional vector sequence in general.

As shown in Figure 4, let the P_IFFT of 100MHz-OFDM be 2048, ΔP be 60KHz, the order p of the vector linear predictor is 5, the vector dimension D is 2, Qb/Sa=4, and VLP-VQ and k-means are obtained. Two-dimensional vector signal quantization schematic diagram of the two methods. Obviously, the dynamic range of the signal processed by the VLP-VQ method is much smaller than that of the original k-means clustering method. At the same time, the closer to the central area of the quantization diagram, the smaller the Voronoi interval division, so under the same Qb/Sa condition, The quantization error of the signal obtained by the VLP-VQ method is much smaller than that of the traditional k-means clustering method.

Step 3: The outputted decimal quantized signal 304 is first converted into a binary bit stream, and then PAM-4 encoding 400 is performed. For digital signals transmitted over short distances (usually within a few tens of kilometers) optical fibers, PAM-4 signals can not only improve the spectral efficiency, but also ensure the performance of the transmitted signals, so PAM-4 signals are generally used in digital mobile fronthaul links. 4 digital signals instead of binary bit streams.

Step 4: The PAM-4 modulated signal is first subjected to electro-optic modulation 400, and then introduced into a single-mode fiber for optical path transmission 500. In the remote radio unit, the optical signal is restored to a PAM-4 electrical signal by a photodetector, and then the PAM The -4 electrical signal is decoded, binary-to-decimal conversion is performed to reconstruct the quantized signal 600 .

Step 5: According to the codebook generated in step 2, the quantized signal is subjected to index-codeword mapping 700 to recover the differential vector quantized signal, and then the differential vector quantized signal is restored according to the prediction coefficient matrix and the optimal codebook generated in step 2. The specific process of the quantized signal of the multi-dimensional vector set is shown in Figure 3 and Step 2.

Step 6: According to the inverse transformation of Step 1, perform demodulation 800 of the multi-dimensional vector set, and reconstruct the quantized signal of the I/Q two-channel sampling signal, which is the quantized value of the recovered OFDM-IFFT sampling point.

In order to verify the superiority of this method, when P_IFFT is 2048, ΔP is 60KHz, dimension D is 2, and the prediction order is 5, we respectively use this method and the vector quantization method based on k-means clustering, based on PCM The scalar quantization method of 25km standard single-mode fiber, 10Gbaud/λ IM/DD system is carried out and compared. In the case of error-free transmission, Figure 5 shows the three sampling methods for signal quantization under the conditions that 100MHz-OFDM 4QAM, 16QAM, 64QAM, and 256QAM correspond to Qb/Sa of 3, 4, 5, and 6, respectively. Error vector magnitude (EVM) test results at the receiver. Obviously, the EVM result obtained by quantization by PCM is the worst; the vector quantization based on k-means clustering algorithm is much lower than the EVM value of the method based on PCM under the same quantization bit; and the method based on vector linear prediction is used for vector quantization. Quantization (VLP-VQ), the EVM value is much smaller than the method based on k-means clustering. For example, in the case of 4, the EVM values measured by the three methods based on PCM, k-means, and VLP-VQ are 12.21%, 7.96%, and 6.43%, respectively. Obviously, the method based on VLP-VQ is better. . Moreover, when the prediction order p increases, the EVM value measured by the receiver will decrease.

Claims (3)

1. A digital mobile forward signal quantization method based on vector linear prediction is characterized by comprising the following steps:

step 1: carrying out IFFT on the OFDM modulation signal to obtain a real part I/Q sampling signal and an imaginary part I/Q sampling signal, and constructing the I/Q sampling signals into a multi-dimensional vector set;

step 2: carrying out differential vector quantization on an input multi-dimensional vector set by using a differential vector quantization method based on vector linear prediction, and outputting a column quantization signal formed by a code word-index mapping sequence;

and step 3: PAM-4 coding is carried out on the quantized signal, and the quantized signal is subjected to electro-optic modulation to become an optical signal;

and 4, step 4: inputting an optical signal into a single mode fiber to be transmitted to a radio frequency remote unit, and recovering a quantized signal after photoelectric detection and PAM-4 decoding;

and 5: performing index-code word mapping and differential demodulation on the recovered quantized signal according to the inverse process of the step 2 to recover an original differential vector quantized signal;

step 6: and (3) demodulating the vector set of the recovered differential vector quantization signals according to the inverse process of the step (1) to recover I/Q two-path sampling signals of the OFDM-IFFT.

2. The vector linear prediction-based digital mobile fronthaul signal quantization method according to claim 1, wherein the differential vector quantization of the input multi-dimensional vector set by using the vector linear prediction-based differential vector quantization method comprises:

step 21: setting parameters of vector linear prediction, including the length of a training vector sequence and a prediction order;

step 22: an offline training process: intercepting a multi-dimensional vector set with a specific length as a training sample sequence, and obtaining a required prediction coefficient matrix and an optimal codebook by using a vector linear prediction method;

step 23: an online quantization process: and inputting a multi-dimensional vector set signal into a differential vector quantizer based on vector linear prediction by using the obtained prediction coefficient matrix and the optimal codebook, wherein the output quantized signal is an index number sequence corresponding to each code word in the optimal codebook, and the vector quantization adopts an algorithm based on k-means clustering.

3. The vector linear prediction-based digital motion forwarding signal quantization method according to claim 2, wherein the specific process of step 22 is:

step 1) intercepting a multidimensional vector set with a certain length of N from the constructed multidimensional vector set as an offline training sequence X ═ s₁,s₂,…,s_N}；

Step 2) through predicting coefficient matrix set A ═ A₁,A₂,…,A_pDefining a finite memory vector linear predictor with a memory length p; a. the_jIs a DxD matrix, j is an element [1, p ]]P is the order of the vector linear predictor; d is the dimension of the multi-dimensional vector set;

step 3) for on-line quantization, the input to the vector quantizer is the predicted vector error expressed as:

wherein s is_nIs a multi-dimensional vector of the input,is the output vector value of the predictor;

p-order vector linear predictor passes through sampling observed value s of previous p times_n-1,s_n-2,…,s_n-pThe input multi-dimensional vector s is obtained by_nPredicted value of (2)

Step 4), defining the performance measure J of the p-order predictor as follows:

selecting a prediction coefficient matrix set A which minimizes the performance measure J of the p-order predictor:

defining a DxD correlation matrix as

Where i, j ∈ [1, p ]]，s_n-i,

Respectively representing D × 1 and 1 × D dimensional vectors;

according to the orthogonality rule, when each error component generated by the linear predictor is orthogonal to each observed value, there is

Substituting the formulas (1) and (2) into the formula (3) to obtain

In which v is [1, p ]]；A_vThe corresponding v-th sub-matrix in the prediction coefficient matrix set in the step 2 is obtained;

writing equation (4) as:

equation (5) is expressed by a matrix equation as follows:

for the formula (6), solving a coefficient matrix set A of the p-order stable D-dimensional vector sequence optimal linear predictor under the general condition by adopting a Levinson-Durbin algorithm;

step 5) clustering the multi-dimensional vector set with the length of N by adopting a k-means clustering method to obtain an optimal codebook C; the length of the codebook is determined by the number of quantization bits Qb ═ log₂And k are the clustering number of k-means.

Images