Method for nonlinear compensation of coherent high-capacity high-order qam system

A method for nonlinear compensation of coherent high-capacity high-order QAM optical communication system, the system comprising an optical phase conjugation (OPC) and a receiver, the method comprising:

• performing, by the OPC, steps of:
• receiving, by the OPC for providing a signal resource to the receiver, a transmitted signal on an intermediate link;
• performing phase conjugation on the transmitted signal to generate an idler, which is an idle light;
• and
• transmitting the idler to the receiver, performing phase recovery on the idler to obtain a constellation diagram, and performing nonlinear compensation on the constellation diagram by using a trained and learned complex-valued deep neural network (CVDNN), and calculating a Q-factor based on the compensated constellation diagram to evaluate communication performance;
• and
• performing, by the receiver, steps of:
• deploying the OPC on the intermediate link of communication between a transmitter and the receiver;
• receiving the idler generated by performing phase conjugation on the transmitted signal based on the OPC;
• performing phase recovery on the idler to obtain the constellation diagram, simulating a nonlinear function relationship between a transmitted signal and a received signal by using the trained and learned CVDNN, and performing nonlinear compensation on the constellation diagram;
• and
• calculating the Q-factor based on the compensated constellation diagram, and evaluating communication performance by using the Q-factor,
• wherein a training and learning process of the CVDNN comprises:
• representing the constellation diagram using a complex value symbol sequence S(n), selecting a training symbol sequence X(n) from the complex value symbol sequence S(n) as an input value of an input layer of the CVDNN, X(n) being denoted as [x(n+k−1), . . . , x(n−1), x(n), x(n+1), . . . , x(n−k+1)], and simultaneously determining that a memory length of the CVDNN is K=2k−1, wherein n is an index of an nth symbol, and k refers to K previous symbols and K next symbols adjacent to the nth symbol;
• and
• transferring the training symbol sequence X(n) step by step from the input layer to an output layer of the CVDNN, and using an output value of a former layer as an input value of a latter layer for training and learning in a transfer process until the output layer outputs a result after training and learning,
• wherein the transferring the training symbol sequence X(n) step by step from the input layer to an output layer of the CVDNN, and using an output value of a former layer as an input value of a latter layer for training and learning in a transfer process until the output layer outputs a result after training and learning comprises:
• transferring the training symbol sequence X(n) of the input layer of the CVDNN to a first hidden layer in a plurality of links to perform training and learning, and calculating an output value of the first hidden layer from the training symbol sequence X(n) on each link by using a training formula of the first hidden layer, the training formula of the first hidden layer being
• h
• 1
• =
• F
• ⁡
• (
• ∑
• i
• =
• 1
• K
• w
• i
• 1
• ⁢
• h
• i
• 1
• +
• b
• i
• 1
• )
• ,
• wherein h1 represents the output value of the first hidden layer, w1 represents a weight value of the first hidden layer, b1 represents a bias vector of the first hidden layer, an activation function F is a Sigmoid nonlinear function of a complex value, and K is the memory length of the CVDNN;
• using the output value h1 of the first hidden layer as an input value of a second hidden layer for training and learning, and calculating an output value of the second hidden layer from the output value h1 of the first hidden layer on each link by using a training formula of the second hidden layer, the training formula of the second hidden layer being
• h
• 2
• =
• F
• ⁡
• (
• ∑
• i
• =
• 1
• M
• w
• i
• 2
• ⁢
• h
• i
• 2
• +
• b
• i
• 2
• )
• ,
• wherein h2 represents the output value of the second hidden layer, w2 represents a weight value of the second hidden layer, b2 represents a bias vector of the second hidden layer, an activation function F is a Sigmoid nonlinear function of a complex value, and M is a quantity of first hidden layers;
• and
• using the output value h2 of the second hidden layer as an input value of the output layer for training and learning, and calculating the output value of the output layer from the output value h2 of the second hidden layer on each link by using a training formula of the output layer, the training formula of the output layer being
• o
• =
• ∑
• i
• =
• 1
• N
• w
• i
• 3
• ⁢
• h
• i
• 2
• +
• b
• i
• 3
• ,
• wherein h3 represents the output value of the output layer, w3 represents a weight value of the output value, b3 represents a bias vector of the output layer, and N is a quantity of second hidden layers.