CN111988249B - A receiving end equalization method based on adaptive neural network and receiving end - Google Patents

Abstract

The invention discloses a receiver equalization method based on an adaptive neural network and a receiver. The method is as follows: 1) Embedding an adaptive neural network in the digital signal processing flow of the receiving end, including the neural network, a decision unit and a loss calculation unit; 2) Using the eigenvectors generated by the received symbols as training data and the corresponding transmitted symbols as labels , train the neural network; 3) use the parameters of the trained neural network as the initialization parameters of the adaptive neural network; 4) input the eigenvector corresponding to the receiving symbol to be equalized into the adaptive neural network, and obtain the corresponding output as y and As an equalized symbol output; 5) y gets a pseudo-label after passing through the decision unit

6) The loss calculation unit calculates y and

The error L between them, and the gradient of L to the neural network parameters; 7) Calculate the average gradient to update the neural network parameters, and perform equalization processing on the subsequent received symbols to be equalized before outputting.

Description

A receiving end equalization method based on adaptive neural network and receiving end

technical field

The invention belongs to the field of optical communication transmission, and relates to a receiving end equalization method, in particular to an adaptive neural network-based receiving end equalization method.

Background technique

Optical fiber communication transmission system has the advantages of large capacity and low cost. Overcoming the linear and nonlinear impairments of optical signals during transmission has become the primary problem for further improving the capacity of optical fiber communication transmission systems. For this problem, there are two types of traditional solutions, one is Time Domain Equalization (TDE), and the other is Frequency Domain Equalization (FDE).

In recent years, due to its powerful fitting ability, neural network has been successfully applied in optical fiber communication transmission system to compensate the linear and nonlinear damage of optical signal during transmission. As a data-driven algorithm, the neural network can compensate the signal not only in the time domain, but also in the frequency domain.

The neural network compensation for signal damage is mainly divided into two stages: training and equalization. In the training phase, the neural network optimizes its own network parameters according to the training sequence; in the equalization phase, the neural network with optimized parameters is used to compensate the signal. Due to the strong fitting ability of neural networks, well-trained neural networks are usually able to compensate the signal well. However, since the parameters of the neural network are not updated once the training is completed, if the damage characteristics of the signal in the training and equalization stages are different, or the transmission channel changes over time, the compensation performance of the neural network will also drop sharply.

Contents of the invention

Aiming at the technical problems existing in the prior art, the object of the present invention is to provide a receiving end equalization method based on an adaptive neural network.

To achieve the above object, the present invention adopts the following technical solutions:

1. The signal itself may be processed by one or more equalization algorithms at the receiving end. Under the premise that the input of the adaptive neural network is symbolic information, the adaptive neural network can be embedded anywhere in the digital signal processing process. After determining the embedding position of the adaptive neural network, the received symbols of the optical fiber communication transmission system are collected and stored there.

2. The stored received symbols are used to generate feature vectors in a specific way as training data, and the corresponding symbols sent by the sender are used as labels to train the neural network. The feature vector corresponding to the kth received symbol is denoted as x _k .

3. Use the parameters of the trained neural network as the initialization parameters of the adaptive neural network. The self-adaptive neural network that the present invention proposes comprises neural network, judgment unit and loss computing unit; Promptly adds the adaptive structure that can realize parameter update on the basis of neural network; Wherein neural network can be artificial neural network (ANN) or other Neural Networks.

4. The eigenvector x corresponding to the received symbol to be equalized is sequentially input into the adaptive neural network, and the corresponding output of the network is the equalized symbol, denoted as y. If the transmitted symbol corresponding to the received symbol is a real number, then y is a real number; if the transmitted symbol is a complex number, then y is a vector, which contains two elements corresponding to the real part and the imaginary part of the transmitted symbol respectively.

判决方式为

其中yⁱ表示发送符号集S中的第i个符号。5. The output y of the adaptive neural network is judged to obtain a pseudo-label

Judgment is

where y ⁱ represents the i-th symbol in the transmitted symbol set S.

以及误差L对网络参数θ的梯度

θ表示神经网络的所有参数。6. Calculate the error between the network output and the pseudo-label using the squared error loss function

And the gradient of the error L to the network parameter θ

θ represents all the parameters of the neural network.

其中B≥1，g_k表示第k个特征向量所对应的梯度。

y_k与

分别表示自适应神经网络的第k个输出及其对应的伪标签。7. Calculate the average gradient corresponding to the B feature vectors

Where B≥1, g _k represents the gradient corresponding to the kth eigenvector.

y _k with

denote the kth output of the adaptive neural network and its corresponding pseudo-label, respectively.

其中η为自适应神经网络的学习率。8. Update network parameters:

where η is the learning rate of the adaptive neural network.

The invention also discloses a receiving terminal based on an adaptive neural network, which is characterized in that it includes an adaptive neural network for equalizing the received symbols in the digital signal processing flow; the adaptive neural network includes a neural network , a decision unit and a loss calculation unit; wherein, the received symbol generates a feature vector as training data, and the sent symbol corresponding to the received symbol is used as a label to train the neural network; and the trained neural network parameters are used as the Describe the initialization parameters of the adaptive neural network;

The neural network is used to equalize the eigenvector x corresponding to the input received symbol, and the corresponding output is recorded as y and output as an equalized symbol; if the transmitted symbol corresponding to the received symbol is a real number, then y is a real number; If the transmitted symbol corresponding to the received symbol is a complex number, then y is a vector, including the real part and the imaginary part of the transmitted symbol;

The judging unit is used for judging y to obtain a pseudo-label

之间的误差L，以及误差L对所述神经网络参数θ的梯度

然后计算B个特征向量对应的平均梯度

然后更新神经网络参数

对后续待均衡的接收符号进行均衡处理后输出，其中，B≥1，g_k表示第k个特征向量所对应的梯度；η为所述自适应神经网络的学习率。The loss calculation unit is used to calculate the network output y and the pseudo-label

The error L between, and the gradient of the error L to the neural network parameter θ

Then calculate the average gradient corresponding to the B feature vectors

Then update the neural network parameters

Perform equalization processing on subsequent received symbols to be equalized and output, wherein, B≥1, g _k represents the gradient corresponding to the kth eigenvector; η is the learning rate of the adaptive neural network.

Compared with prior art, positive effect of the present invention is:

Compared with the traditional neural network equalizer, the adaptive neural network equalizer has higher equalization performance, and the specific experimental results are shown in Figure 4.

Description of drawings

Fig. 1 is a schematic structural diagram of a training phase of an optical fiber communication transmission system according to an embodiment of the present invention.

Fig. 2 is a schematic structural diagram of an adaptive neural network in an embodiment of the present invention.

Fig. 3 is a schematic structural diagram of an equalization stage of an optical fiber communication transmission system according to an embodiment of the present invention.

Fig. 4 is a schematic diagram of a 960km transmission experiment result of a WDM system with 64Gbaud and 16QAM signals according to an embodiment of the present invention.

detailed description

The present invention will be described in further detail below through specific embodiments and accompanying drawings.

1. The optical signal at the receiving end undergoes front-end digital processing (dispersion compensation, phase recovery, etc.) to obtain received symbols, and generate corresponding eigenvectors according to the received symbols. The eigenvector x _k of the kth received symbol s _k consists of itself and 2L adjacent received symbols before and after it, that is, x _k =[Re(s _kL ), Im(s _kL ),...,Re(s _k ),Im(s _k ),...,Re(s _k+L ),Im(s _k+L )]. Re(s _k ) represents the real part of the complex symbol _sk , and Im(s _k ) represents the imaginary part of the complex symbol s _k . The value of L needs to be optimized according to the actual system. Generally, the higher the transmission rate of the system and the longer the transmission distance, the larger the value of L.

2. The generated N _tr feature vectors are used as training data, and the corresponding sent symbols are used as labels, and a neural network is trained using the square error loss function. The above process is shown in Figure 1. If the transmitted symbol is a real number, the symbol itself is used as a label; if the transmitted symbol is a complex number, a one-dimensional vector composed of the real part and the imaginary part of the complex signal is used as a label.

判决规则如下：3. Use the parameters of the trained neural network as the initialization parameters of the adaptive neural network. The specific structure of the adaptive neural network is shown in Figure 2. The feature vector x _j generated by the jth symbol s _j to be equalized is used as the input of the adaptive neural network, and its output is denoted as y _j . After the output y _j is judged, the pseudo-label is obtained

Judgment rules are as follows:

之间的误差

以及L对网络参数θ的梯度

Wherein y ⁱ represents the i-th symbol in the transmitted symbol set S, taking a 16QAM signal as an example, the transmitted symbol set consists of 16 complex symbols. Calculate the network output y _k and the pseudo-label using the squared error loss function

error between

and the gradient of L to the network parameter θ

4. Calculate the average gradient corresponding to consecutive B feature vectors:

where B≥1. Update the parameters of the adaptive neural network based on the computed average gradient:

where η is the learning rate of the adaptive neural network. It is worth noting that when the adaptive neural network is used to equalize signal impairments, the network parameters are updated every B symbols equalized.

5. Use the adaptive neural network after parameter update to continue to equalize subsequent data. The above process is shown in Figure 3.

Fig. 4 is a schematic diagram of a 960km experimental result of a WDM system using 64Gbaud and 16QAM signals according to an embodiment of the present invention. The horizontal axis is the fiber input power, and the vertical axis is the signal Q ² value. From the equalization results of the adaptive neural network and the traditional neural network, it can be seen that the adaptive neural network has higher equalization performance, because the adaptive neural network can dynamically adjust its own network parameters according to the damage characteristics of the signal.

The above embodiments are only used to illustrate the technical solution of the present invention and not to limit it. Those of ordinary skill in the art can modify or equivalently replace the technical solution of the present invention without departing from the spirit and scope of the present invention. The scope of protection should be determined by the claims.

Claims (10)

1. A receiving end equalization method based on an adaptive neural network comprises the following steps:

1) Embedding a self-adaptive neural network in a digital signal processing flow of a receiving end, wherein the self-adaptive neural network is used for balancing a received symbol; the self-adaptive neural network comprises a neural network, a judgment unit and a loss calculation unit;

2) Training the neural network by using the characteristic vector generated by the received symbol as training data and using a transmitting symbol corresponding to the received symbol as a label; wherein the eigenvector corresponding to the kth received symbol is marked as x _k ；

3) Taking the parameters of the trained neural network as initialization parameters of the adaptive neural network;

4) Sequentially inputting the characteristic vectors x corresponding to the received symbols to be equalized into the adaptive neural network to obtain corresponding outputs which are recorded as y and serve as equalized symbol outputs; if the sending symbol corresponding to the receiving symbol is a real number, y is a real number; if the transmitting symbol corresponding to the receiving symbol is a complex number, y is a vector including a real part and an imaginary part of the transmitting symbol;

5) y is processed by the judgment unit to obtain a pseudo label

6) Loss calculation unit calculates network output y and pseudo label

Error L between, and gradient of error L to the neural network parameter θ

7) Calculating the average gradient corresponding to the B feature vectors

Wherein B is not less than 1,g _k Representing the gradient corresponding to the k-th feature vector; then updating the neural network parameters

And carrying out equalization processing on the subsequent receiving symbols to be equalized and then outputting the symbols, wherein eta is the learning rate of the adaptive neural network.

2. The method of claim 1, wherein the loss computation unit computes the net output y and the pseudo label using a squared error loss function

Error between

3. The method of claim 1, wherein the kth received symbol s _k Corresponding feature vector x _k Is composed of itself and 2L adjacent received symbols before and after it.

4. The method of claim 3, wherein the feature vector x _k ＝[Re(s _k-L ),Im(s _k-L ),...,Re(s _k ),Im(s _k ),...,Re(s _k+L ),Im(s _k+L )](ii) a Wherein, re(s) _k ) Representing a complex symbol s _k Real part of, im(s) _k ) Representing complex symbols s _k The imaginary part of (c).

5. The method of claim 1, wherein the decision unit decides in such a way that

Wherein y is ⁱ Indicating the ith symbol in the transmit symbol set S.

6. A receiving end based on a self-adaptive neural network is characterized by comprising the self-adaptive neural network, a receiving end and a processing device, wherein the self-adaptive neural network is used for equalizing a received symbol in a digital signal processing flow; the self-adaptive neural network comprises a neural network, a judgment unit and a loss calculation unit; training the neural network by using the characteristic vector generated by the received symbol as training data and using a transmitting symbol corresponding to the received symbol as a label; taking the parameters of the trained neural network as the initialization parameters of the adaptive neural network;

the neural network is used for balancing a characteristic vector x corresponding to an input receiving symbol to obtain a corresponding output which is recorded as y and is used as a balanced symbol output; if the sending symbol corresponding to the receiving symbol is a real number, y is a real number; if the transmitting symbol corresponding to the receiving symbol is a complex number, y is a vector including a real part and an imaginary part of the transmitting symbol;

the judgment unit is used for judging y to obtain a pseudo label

The loss calculation unit is used for calculating the network output y and the pseudo label

Error L between, and gradient of error L to the neural network parameter θ

Then calculating the average gradient corresponding to the B feature vectors

Then updating the neural network parameters

Equalizing and outputting subsequent received symbols to be equalized, wherein B is more than or equal to 1,g _k Representing the gradient corresponding to the k-th feature vector; eta is the learning rate of the adaptive neural network.

7. The receiving end according to claim 6, characterized in that the kth received symbol s _k Corresponding feature vector x _k Is composed of itself and 2L adjacent received symbols before and after it.

8. The receiving end of claim 7, wherein the eigenvector x is characterized by _k ＝[Re(s _k-L ),Im(s _k-L ),...,Re(s _k ),Im(s _k ),...,Re(s _k+L ),Im(s _k+L )](ii) a Wherein, re(s) _k ) Representing a complex symbol s _k Real part of, im(s) _k ) Representing complex symbols s _k The imaginary part of (c).

9. The receiving end of claim 6, wherein the loss calculating unit calculates the network output y and the pseudo label using a squared error loss function

Error therebetween

10. The receiving end of claim 6, wherein the decision unit decides in a manner of

Wherein y is ⁱ Indicating the ith symbol in the transmit symbol set S.

Images