CN108494488B

CN108494488B - DFE-based SVM (support vector machine) equalization method for short-distance optical communication system

Info

Publication number: CN108494488B
Application number: CN201810145616.6A
Authority: CN
Inventors: 毕美华; 卓先好; 姜伟; 俞嘉生; 杨国伟; 周雪芳; 胡淼; 骆懿; 李齐良
Original assignee: Hangzhou Dianzi University
Current assignee: Hangzhou Dianzi University
Priority date: 2018-02-12
Filing date: 2018-02-12
Publication date: 2021-04-30
Anticipated expiration: 2038-02-12
Also published as: CN108494488A

Abstract

The DFE-based SVM equalization method used in the short-distance optical communication system of the present invention includes: Step 1: a digital signal with a training sequence is converted into an optical signal through electro-optical modulation, and is converted into an electrical signal at the receiving end and extracted by sampling Training sequence; Step 2: Construct the eigenvector of the training symbol according to the DFE structure, and use the SVM to calculate the optimal hyperplane based on the eigenvalue vector of the training sequence; Step 3: Using the hyperplane, the input information sequence based on the DFE structure Determine to achieve equalization, and then restore the original signal through demodulation; Step 4: Save the symbols with the set length after equalization, and become the eigenvalue of the next symbol to be detected by the preamble interference through feedback input; Step 5: Information sequence every Return to step 2 for a certain length and retrain the optimal hyperplane. The invention solves the problem that the system receiving sensitivity decreases due to the dispersion caused by the high-speed transmission of the signal in the optical fiber and the randomly distributed Gaussian noise caused by the system device.

Description

DFE-based SVM (support vector machine) equalization method for short-distance optical communication system

Technical Field

The invention relates to the technical field of optical transmission, in particular to a precision feedback equalization (Decision feedback equalization) -based SVM (support vector machine) equalization method for a short-distance optical communication system.

Background

With the rapid development of new services such as cloud computing, high-definition video, virtual reality and the like, the demand of terminal users on bandwidth is increasing, long-distance optical fiber transmission technology is mature day by day at present, large-capacity data transmission can be realized, the performance is relatively stable, but the bandwidth of an access network of the long-distance optical fiber transmission technology still cannot meet the demand of the users, and therefore the problem of short-distance optical transmission is solved. Because the short-distance optical transmission technology is fast to upgrade, large in scale and high in investment, high performance and low cost are key factors for determining the evolution of the short-distance optical transmission technology. At present, the single-channel data capacity of optical communication reaches 100Gb/s, but an industrial chain of high-speed optical devices is not formed yet, the cost is high, and the method is not suitable for large-area popularization. When modulation transmission is performed using an optical device with a low bandwidth, Inter Symbol Interference (ISI) is caused by problems such as dispersion and bandwidth limitation. Therefore, in order to meet the requirements of high performance and low cost of optical transmission, an equalization technique can be used to counteract the intersymbol interference caused by channel fading.

The existing literature search shows that the short-distance Optical transmission mainly eliminates interference on an Optical domain or an electrical domain of a receiving end at present, for example, Lei Xue, et al, published symmetry 100-Gb/s TWDM-PON in O-band based on10G-Class Optical Devices Enabled by Dispersion compensation, and utilizes a delay interferometer to shape and compensate a spectrum to realize the balance on the Optical domain. And then as in Junqi xia et al, "investment on adaptive equalization techniques for 10G-glass optics based 100G-PON system", on the electric domain of the receiving end of the short-distance high-speed transmission system, the signal is compensated by equalization techniques such as DFE (decision feedback equalization), FFE (feed forward equalization) and the like, so that the sensitivity of the receiving end is improved to a certain extent, the feasibility of the scheme is verified, the cost is low, and the short-distance low-cost high-speed transmission principle is met. However, the training sequences of the conventional FFE and DFE are long, which increases the channel cost to a certain extent, and cannot effectively compensate the signal when the channel attenuation is severe.

Disclosure of Invention

Aiming at the defects in the prior art, the invention provides a DFE-based SVM equalizing method for a short-distance optical communication system.

The invention adopts the following technical scheme:

a DFE-based SVM equalization method for a short-range optical communication system, comprising the steps of:

step 1: the digital signal with training sequence is converted into optical signal through electro-optical modulation and then sent, and after the optical signal is converted into electric signal at the receiving end, the training sequence is extracted through sampling.

Step 2: and constructing a feature vector of a training code element according to the DFE structure, and calculating an optimal hyperplane by using the SVM based on the feature value vector of the training sequence.

And step 3: and (3) utilizing the hyperplane, judging the input information sequence based on the DFE structure to realize equalization, and then recovering the original signal through demodulation.

And 4, step 4: and storing the equalized code element with a certain length, and inputting the code element after feedback to form a characteristic value of the next code element to be detected subjected to preamble interference.

And 5: and returning the information sequence to the step 2 at regular intervals, and retraining the optimal hyperplane.

Preferably, in step 1, the training sequence is a pseudo-random sequence, and the information sequence is inserted into a series of training sequences at intervals, where the length of the intervals is set according to the degree of influence of the time-varying characteristic of the channel.

Preferably, in step 2, the feature vector is constructed by: the DFE structure is divided into a feedforward part and a feedback part, so that a plurality of front code elements of the training code elements to be detected are taken as a feedforward part, correct decisions of a plurality of rear code elements of the training code elements to be detected are taken as a feedback part, then the feedforward and feedback parts are combined to be taken as a characteristic value vector of the current training code element, and then the characteristic value vectors of n-k training code elements are stored (wherein n is the number of all training code elements, and k is the number of tap delayers of a feedback filter).

Preferably, in step 2, the SVM calculates an optimal hyperplane: and (3) the constructed training sequence characteristic value vector and a training sequence (the latter is the correct judgment result of the former) regenerated by the receiving end enter the SVM together for training, and the optimal hyperplane is calculated.

The specific steps of SVM (support vector machine) to calculate the optimal hyperplane are as follows:

1) initializing the hyperplane, calculating the distance from each feature vector point to the hyperplane, and taking the feature vector point closest to the hyperplane as a support vector.

2) To maximize the robustness of the hyperplane, the hyperplane needs to be adjusted to maximize the separation of the support vector from the hyperplane.

3) And equivalently converting the interval maximization in the step 2 into minimization, thereby meeting the solving of the convex optimization problem.

4) And introducing Lagrange number multiplication to solve the conditional extremum of the convex optimization problem to obtain the corresponding relation between the hyperplane normal vector, the intercept and the Lagrange multiplier.

5) Substituting the normal vector and intercept represented by the Lagrange multiplier into the original expression, and solving the Lagrange multiplier according to the dual problem and the SMO algorithm so as to obtain the optimal hyperplane.

Preferably, in step 3, based on the DFE structure, the normal vector of the optimal hyperplane is regarded as all tap coefficient sets in the feedforward and feedback filters, the information sequence is multiplied by the normal vector of the optimal hyperplane through the tap delayer to obtain the weighted sum of the feedforward part and the feedback part, the equalized signal is obtained through the judgment, and then the original signal is restored through the demodulation.

Preferably, in step 4, the equalized symbol with a certain length is stored: the intersymbol interference mainly comes from trailing interference of a front code element sequence and leading interference of a rear code element sequence, so that the length of the front code element sequence and the rear code element sequence is similar to the length of the intersymbol interference to achieve a good balance effect. Based on DFE structure, equalized signal is post code element sequence, so that code element with corresponding length of equalized output signal is stored according to length of post code element sequence of interference.

Preferably, in step 4, based on the DFE structure, the symbol outputted from the equalization enters the feedback filter and is stored by the tap delay, so as to be used as the characteristic value of the preamble interference of the next symbol to be detected.

Preferably, in step 5, since most channels have time-varying characteristics, the hyperplane needs to be retrained every certain information sequence length to ensure the equalization effect.

The SVM is a machine method based on a statistical learning theory, and is equivalent to a maximum interval classifier. The classification of the data is realized by inputting the feature value of each sample and finding the maximum interval linear classifier (optimal hyperplane) in the feature space. And the SVM has strong robustness, and the optimal hyperplane conforming to the sample characteristic distribution can be trained only by small samples, so that classification is realized. Equalization is to offset intersymbol interference by continuously adjusting the tap coefficient of the equalizer through a self-adaptive algorithm, so that the equalization output is continuously close to the initial signal, and thus correct decision classification is realized. And the optimal hyperplane model of the SVM is similar to the equalizer model, so the classification of the SVM can be applied to the equalization to improve the performance.

The equalization method of the present invention is based on an improved version of the DFE SVM. Wherein the DFE removes trailing interference from a previous symbol to a current symbol by a feedforward filter and removes leading interference caused by a previously detected symbol by a feedback filter, so that when an information symbol is detected and determined, interference from the current symbol to a subsequent symbol is removed before the subsequent symbol is detected. To take advantage of the high efficiency of SMO (sequence minimization) algorithm computation in SVMs, the adaptive algorithm of the DFE is replaced with SMO. Therefore, in the DFE-based SVM, a front code element and a code element (a rear code element) output by equalization are taken as characteristic values of intersymbol interference of a current code element to be detected, so that a characteristic vector is constructed, all the characteristic vectors are input into an SMO algorithm to obtain an optimal hyperplane, then the hyperplane is used for judging an input information sequence to realize equalization, and finally an original signal is recovered through demodulation. In consideration of the time-varying characteristic of a channel, a training sequence is inserted into an information sequence at intervals of a certain length, and the optimal hyperplane balancing ensuring effect is obtained through retraining. And the SVM fully utilizes the high efficiency of the SMO algorithm, and compared with the traditional DFE self-adaptive algorithm, the equalization of the channel can be completed only by a short training sequence, so that the equalization can be still completed quickly by calculating the optimal hyperplane for multiple times by the SVM.

Compared with the prior art, the invention has the following beneficial effects:

1. compared with the traditional equalization technology, the equalization technology of the invention reduces the error rate and improves the precision under the training sequence with the same length.

2. The error rate of the equalization technology of the invention is less influenced by the length of the training sequence, the short training sequence can also ensure the low error rate, and the utilization rate of the frequency band is improved.

3. The equalization technology of the invention has high adaptability to the channel and can adapt to the channels with different characteristics.

The invention relates to a DFE-based SVM equalization method for a short-distance optical communication system, which comprises the steps of firstly extracting a training sequence in a signal at a receiving end, then constructing a feature vector of a code element subjected to leading and trailing interference based on a DFE structure, calculating an optimal hyperplane through training of an SVM, and finally judging an information sequence input based on the DFE structure by utilizing the optimal hyperplane to realize equalization. The invention solves the problem that the receiving sensitivity of the system is reduced due to dispersion caused by high-speed transmission of signals in the optical fiber and randomly distributed Gaussian noise brought by system devices.

Drawings

Other features, objects, and advantages of the invention will become apparent by comparison to other equalization techniques upon reading the following drawings:

fig. 1 is a schematic structural diagram of a short-distance high-speed optical transmission system.

Fig. 2 is a schematic diagram of a conventional decision feedback equalization principle.

Fig. 3 is a schematic diagram of the principle of SVM-based decision feedback equalization.

Fig. 4 is a graph showing BER performance comparison of signals compensated by different equalization methods at the receiving end after the signals pass through the short-distance optical transmission system, where: the horizontal axis is the optical power of the receiving end, the vertical axis is the BER which represents the bit error rate, wherein the comparison of the BER performance of the receiving end respectively comprises: DFE with RLS as adaptive algorithm, FFE with RLS as adaptive algorithm and decision feedback equalization based on SVM without equalization.

Detailed Description

The present invention will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the invention, but are not intended to limit the invention in any way. It should be noted that it would be obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit of the invention. All falling within the scope of the present invention. For this reason, the present invention performs experiments using 0,1 signals to verify the effect of equalization.

The invention relates to a DFE-based SVM equalization method for a short-distance optical communication system, which comprises the steps of firstly carrying out high-speed electro-optical modulation on a digital signal with a training sequence, then transmitting the digital signal into a standard single-mode optical fiber, converting the optical signal into an electric signal through an optical detector at a receiving end, then sampling the electric signal to extract a corresponding training sequence, constructing a feature vector of a training code element subjected to intersymbol interference according to a DFE structure, then training the feature vector of the training sequence by using an SVM to calculate an optimal hyperplane, then judging an information sequence by using the hyperplane based on the DFE structure to obtain an equalized signal, recovering an original signal through demodulation, and simultaneously, enabling the equalized output signal to enter a tap delayer in a feedback filter to be stored as a feature value of leading interference of a next symbol to be detected. The method specifically comprises the following steps:

step 1: the 0,1 signal with the training sequence is converted into an optical signal through high-speed electro-optical modulation.

Step 2: the optical signal is transmitted by a single mode fiber, and after being converted into an electric signal at a receiving end, the electric signal is subjected to up-sampling, timing recovery and down-sampling, and a training sequence is extracted.

And step 3: and constructing a feature vector of the training code element according to the DFE structure, and calculating an optimal hyperplane by using the SVM based on the feature value vector of the training sequence.

And 4, step 4: and (3) utilizing the hyperplane, judging the input information sequence based on the DFE structure to realize equalization, and then recovering the original signal through demodulation.

And 5: and storing the equalized code element with a certain length, and inputting the code element after feedback to form a characteristic value of the next code element to be detected subjected to preamble interference.

Step 6: and (3) returning the information sequence to the step 3 at regular intervals, and retraining the optimal hyperplane.

The terms in the steps are further explained:

1. a conventional DFE consists of two parts, a feedforward part and a feedback part. As shown in fig. 2, the outputs are as follows:

wherein x is an input signal and x is an output signal,

for decision output, n, w are the tap number and coefficient of the feedforward filter, respectively, and m, b are the tap number and coefficient of the feedback filter, respectively. In the conventional decision feedback equalization, the tap coefficient of the equalizer is continuously adjusted by an adaptive algorithm through an input training sequence, so that the frequency characteristic of a channel is estimated and a signal is compensated. Wherein the signal is feed forward equalized to eliminate the trailing interference of the preceding symbol, and the feedback part is used to eliminate the leading interference of the following symbol. The DFE-based SVM of the present invention adjusts the weight of each eigenvalue using the SMO algorithm instead of the adaptive algorithm.

2. Construction of eigenvalue vectors of symbols in DFE structure

As known from 1, the input of decision feedback equalization consists of two parts, one part being the signal input part and the other part being the feedback part of the output signal. The feature vector of the ith training symbol based on the decision feedback equalization structure can be expressed as:

wherein

Is x_iAnd (5) outputting the judged output value.

3. SVM (support vector machine)

The SVM is a maximum interval classifier, that is, an n-dimensional hyperplane is searched in sample data to divide the data into two categories, so that the distance between the two categories is maximum.

The SVM calculates the optimal hyperplane according to the following specific steps:

2) To maximize the robustness of the hyperplane, the hyperplane needs to be adjusted to maximize the sum of the support vectors to the hyperplane.

4) And introducing Lagrange number multiplication to solve the conditional extremum of the convex optimization problem to obtain the corresponding relation between the hyperplane method vector and the intercept and the Lagrange multiplier.

Because the traditional channel model is similar to the hyperplane equation generated by the SVM, the hyperplane equation can be considered to approximate the channel model, then the classification capability of the hyperplane in the SVM is utilized to realize the compensation of the signal, and the judgment result is the balanced output after judgment.

And (3) channel model:

hyperplane equation: w^TX+b＝0

Step 1: the insertion mode of the training sequence is to insert the training sequence in the data signal at regular intervals for synchronization and channel estimation of the receiving end.

The training sequence in step 1 is a pseudo random 0,1 sequence generated by a sequence generator.

The optical modulation in the step 1 adopts an external modulation mode, a modulation signal controls an external modulator connected behind a laser, and the intensity of output light of the modulator is changed along with the signal by utilizing the physical effects of electro-optic, acousto-optic and the like of the modulator.

Step 2: the optical signal is converted into an electrical signal by a photodetector, which employs a photodiode.

Step 2: the method comprises the steps of sampling signals by using a high-speed storage oscilloscope, firstly carrying out upsampling by using a sampling rate 32 times of a data rate, then timing recovery and extraction of a clock signal, synchronizing a digital sequence, and then carrying out downsampling at the data rate to extract a training sequence.

And (3) constructing a feature vector: the DFE structure is divided into a feedforward part and a feedback part, so that a plurality of front code elements of the training code elements to be detected are taken as a feedforward part, correct decisions of a plurality of rear code elements of the training code elements to be detected are taken as a feedback part, then the feedforward and feedback parts are combined to be taken as a characteristic value vector of the current training code element, and then the characteristic value vectors of n-k training code elements are stored (wherein n is the number of all training code elements, and k is the number of tap delayers of a feedback filter).

In step 3, the SVM calculates an optimal hyperplane: and (3) the constructed training sequence characteristic value vector and a training sequence (the latter is the correct judgment result of the former) regenerated by the receiving end enter the SVM together for training, and the optimal hyperplane is calculated.

And 4, step 4: based on the DFE structure, the normal vector of the optimal hyperplane is regarded as all tap coefficient sets in a feedforward filter and a feedback filter, the information sequence is multiplied by the normal vector of the hyperplane through a tap delayer to obtain the weighted sum of a feedforward part and a feedback part, if the weighted sum is less than 0, the weighted sum is judged to be 0, otherwise, the weighted sum is judged to be 1, the judged result is balanced output, and then the original signal is recovered through demodulation.

And 5, storing the equalized code elements with a certain length: the intersymbol interference mainly comes from trailing interference of a front code element sequence and leading interference of a rear code element sequence, so that the length of the front code element sequence and the rear code element sequence is similar to the length of the intersymbol interference to achieve a good balance effect. Based on DFE structure, the equalized signal is post-code element sequence, so that the code element with corresponding length of the equalized output signal is stored according to the length of post-code element sequence of interference.

And 5: based on the DFE structure, the code element output by equalization enters a feedback filter and is stored by a tap delayer to be used as a characteristic value of leading interference of the next code element to be detected.

Step 6: since most channels have time-varying characteristics, the optimal hyperplane needs to be retrained every certain information sequence length to ensure equalization.

As shown in fig. 3, the DFE based on SVM mainly includes the following steps:

based on DFE structure, input training sequence, take feedforward input sequence x and feedback decision output sequence

And constructing a characteristic value vector of the training sequence, and training the characteristic vector of the training sequence by using the SVM to calculate the optimal hyperplane. And judging the input information sequence by utilizing the hyperplane and based on the DFE structure to obtain an equalized output signal, and simultaneously entering a tap delayer of a feedback filter for storage to be used as a characteristic value of leading interference of a next detection code element. Compared with the traditional DFE, the DFE has a simple structure, does not need a plurality of tap coefficients to compensate signals, can directly judge input signals after a training sequence is only needed to determine the hyperplane, and reduces the circuit complexity.

The length of training sequence for verifying the performance of SVM-DFE is 500, the length interval of the inserted training sequence is 40000, the total length of information sequence is 125000, and three training sequences are inserted in total. The equalization algorithm used by the DFE and FFE to be compared is RLS (least squares), the training sequence length being 2000.

Fig. 4 is a graph showing BER performance comparison of a modulation device with 10G bandwidth generating a signal with 25G rate, transmitted through a 20km standard single-mode fiber, and compensated by different equalization methods at a receiving end, where: with the horizontal axis being the receiving endThe optical power, the vertical axis is BER representing the bit error rate, and 1 × 10 is taken^-3As the BER sensitivity of the receiver. Through comparison, the decision feedback equalization based on the SVM is obviously superior to the traditional DFE and FFE, and can well compensate signals under the condition of low received optical power.

In conclusion, the SVM based on the DFE can well solve the problem of reduced receiving sensitivity caused by symbol distortion, intersymbol interference and the like generated by a low-bandwidth device for transmitting a high-speed signal. Compared with the traditional equalization technology, the method obviously improves the receiving sensitivity, can still effectively compensate the received signal under the condition of lower received optical power, needs less training sequences and reduces the expenditure of the channel to a certain extent. Therefore, the SVM based on the DFE can be better applied to a short-distance optical transmission system and meets the requirements of low cost and high capacity transmission.

The foregoing description of specific embodiments of the present invention has been presented. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention. The embodiments and features of the embodiments of the present application may be combined with each other arbitrarily without conflict.

Claims

1. A DFE-based SVM equalization method for a short-range optical communication system, comprising the steps of:

step 1: the method comprises the steps that a digital signal with a training sequence is converted into an optical signal through electro-optical modulation and sent, the optical signal is converted into an electric signal at a receiving end, then the electric signal is sampled to extract the training sequence, the training sequence is a pseudo-random sequence, and a string of training sequences are inserted into an information sequence at intervals of a certain length;

step 2: constructing a feature vector of a training code element according to a DFE (decision feedback equalization) structure, and calculating an optimal hyperplane by utilizing an SVM (support vector machine) based on the feature value vector of a training sequence; constructing a feature vector: the DFE structure is divided into a feedforward part and a feedback part, a plurality of front code elements of the training code elements to be detected are taken as a feedforward part, correct decisions of a plurality of rear code elements of the training code elements to be detected are taken as a feedback part, then the feedforward and feedback parts are combined to be taken as a characteristic vector of the current training code element, and the characteristic vectors of n-k training code elements are stored, wherein n is the number of all training code elements, and k is the number of tap delayers of a feedback filter;

and step 3: utilizing the hyperplane, judging the input information sequence to realize equalization based on a DFE structure, and then restoring an original signal through demodulation;

and 4, step 4: storing the code element with the set length after equalization, and inputting the code element into a characteristic value of the next code element to be detected subjected to preamble interference through feedback;

2. A DFE-based SVM equalization method for a short distance optical communication system as claimed in claim 1, wherein in step 2, the SVM calculates an optimal hyperplane process: and (3) the constructed training sequence feature vector and a training sequence regenerated by the receiving end enter the SVM together for training, and the optimal hyperplane is calculated.

3. A DFE-based SVM equalization method for short distance optical communication system as claimed in claim 1, wherein in step 3, based on DFE structure, the normal vector of the optimal hyperplane is regarded as all the tap coefficient sets in the feedforward and feedback filters, the information sequence is multiplied by the normal vector of the hyperplane through the tap delayer to obtain the weighted sum of the feedforward part and the feedback part, the equalized signal is obtained through the decision, and then the original signal is recovered through demodulation.

4. A DFE-based SVM equalization method for a short distance optical communication system as claimed in claim 1, wherein in step 4, symbols of a set length after equalization are saved: the intersymbol interference mainly comes from trailing interference of a front code element sequence and leading interference of a rear code element sequence, and the length of the front code element sequence and the rear code element sequence is similar to that of the intersymbol interference so as to achieve a better balance effect; based on DFE structure, equalized signal is post code element sequence, and code element of corresponding length of equalized output signal is stored according to length of post code element sequence of interference.

5. A DFE-based SVM equalization method for short distance optical communication system as claimed in claim 1 or 4, characterized in that in step 4, based on the DFE structure, the symbol outputted from the equalization enters the feedback filter and is stored through the tap delay as the characteristic value of the leading interference of the next symbol to be detected.