Optical fiber nonlinear equalization method based on TL-CLDNN algorithm

Abstract

The invention discloses an optical fiber nonlinear equalization method based on a TL-CLDNN algorithm, and the method comprises the steps: firstly constructing a single-channel DP-16QAM coherent optical fiber communication system, and calculating a simplified triple through employing triple index scaling, calculation value sharing and a symbol degradation algorithm after the carrier phase of a receiving end of the system is recovered; the method comprises the following steps: firstly, obtaining a triple feature and a label at a current moment, then inputting the triple feature and the label at the current moment into a CLDNN network, obtaining an estimated optical fiber nonlinear damage value after the CLDNN network is balanced, applying a part of parameters trained in a source domain to a target domain by adopting transfer learning (TL) based on a multi-source domain, and rapidly reconstructing a target network; and finally, compensating the nonlinear damage of the coherent optical fiber communication system by using the value. According to the method, the calculation of the triad is simplified by utilizing the triad index scaling, calculation value sharing and symbol degradation algorithms, and a part of parameters trained in the source domain are applied to the target domain through transfer learning, so that the calculation complexity of the experiment is reduced. According to the scheme, high system transmission performance can be obtained under low calculation complexity, and the compensation scheme is high in universality.

Description

Technical Field

[0001] The present invention relates to the technical field of optical fiber communication, and particularly to an optical fiber nonlinear equalization method based on the TL-CLDNN algorithm. Background Art

[0002] Optical communication is the cornerstone of modern social information technology. In recent years, with the rapid growth of bandwidth-intensive services such as high-definition video streams, the Internet of Things, and big data, optical networks are developing towards being dynamic and complex. However, for high-speed long-haul optical fiber transmission systems, optical signals are vulnerable to various transmission impairments. Among them, optical fiber linear impairments and nonlinear impairments (NLI) have long been the basic bottlenecks of long-haul optical transmission systems. Fortunately, due to the rapid development of DSP technology in the past few years, linear impairments can be fully compensated. Therefore, NLI has become the main limitation of long-haul optical fiber transmission systems.

[0003] Regarding the nonlinear impairments of optical fibers, although the DBP algorithm can effectively suppress optical fiber nonlinearity, it requires multiple FFT and IFFT processes, which increases the overhead of the system computing unit and the real-time effect is also poor; the compensation performance of the VSNE algorithm is affected by the tap coefficients and is unstable, and the overhead of the system computing unit is also large; the compensation performance of the PCTM algorithm will be at the cost of sacrificing 50% of the spectral efficiency; due to its single-stage and symbol rate processing, the PNC algorithm is regarded as a powerful and potential method. However, in practical applications, it involves a large number of multiplication calculations of perturbation terms. Especially in large dispersion links, it will generate a large overhead of the system computing unit.

[0004] In recent years, machine learning technology has been widely used in various aspects of optical communication due to its powerful self-learning ability. Among them, deep learning algorithms are used for NLI compensation (NLC) in coherent optical fiber transmission systems and have obtained good performance. Based on post-perturbation equalization and a deep neural network (DNN), the DNN-NLC algorithm was proposed. However, the DNN-NLC algorithm needs to calculate perturbation coefficients to select the triple with the greatest contribution. As a feedforward neural network, DNN is a memoryless deep learning algorithm. To meet the feature requirements of the input layer, DNN needs to increase the dimension of the input layer, resulting in higher computational complexity. Summary of the Invention

[0005] Aiming at the above problems, the purpose of the present invention is to provide an optical fiber nonlinear equalization method based on the TL-CLDNN algorithm. The optical fiber nonlinear impairments are equalized by the TL-CLDNN network to obtain the estimated optical fiber nonlinear impairment value, and this value is used to compensate the nonlinear impairments of the coherent optical fiber communication system. It can be independent of the transmission link and can work without prior knowledge of the link parameters. The technical solution is as follows:

[0006] A fiber optic nonlinear equalization method based on the TL..CLDNN algorithm, comprising the following steps:

[0007] Step 1: Calculate the product of the first two symbols in each triple derived from the first-order perturbation solution of the nonlinear Schrödinger equation, and the result is shared by multiple consecutive symbols;

[0008] Step 2: Degrade the remaining symbols in the triple and replace the second multiplication with a simple logical operation;

[0009] Step 3: Cascade the CNN, LSTM, and DNN networks to construct a unified CLDNN network, and equalize the simplified triple to obtain fiber optic nonlinear impairments, thereby compensating the coherent fiber optic communication system;

[0010] Step 4: Adopt transfer learning based on multi-source domains, apply a part of the parameters trained in the source domain to the target domain, and quickly compensate the target coherent fiber optic communication system.

[0011] Further, the calculation and sharing of the product of the first two symbols in the triple in Step 1 include:

[0012] Based on the first-order perturbation theory, the in-channel fiber optic nonlinear distortion of the dual-polarization symbol is approximately expressed as:

[0013]

[0014] where Δu k，x and △u k，y are respectively the fiber optic nonlinear perturbations on the x polarization and y polarization at the k-th moment; P0 is the transmitted power; X k+m and Y k+m respectively represent the transmission symbols on the x polarization and y polarization at the (k + m)-th moment; and respectively represent the conjugates of the transmission symbols on the x polarization and y polarization at the (k + m + n)-th moment; m and n are symbol indices; C m,n represents the nonlinear coefficient of the fiber optic; at this time, the calculation formula of the triple is:

[0015]

[0016] where T k，x and T k,y respectively represent the triples on the x polarization and y polarization at the k-th moment. At this time, calculate X k+n X k+m 、Y k+n Y k+m 、Y k+n X k+m and X k+n Yk+m value, and share the calculated value to other moments.

[0017] Furthermore, the specific steps of step 2 include:

[0018] Degenerate the symbols and into QPSK signals or the origin, and replace X through simple logical operations k+ xX k+m and Y k+n Y k+m and Y k+n X k+m and X k+n Y k+m and and for calculation. At this time, the simplified triple is:

[0019]

[0020] In the formula, Log(·) represents logical operation.

[0021] The beneficial effects of the present invention are: The present invention simplifies the calculation of the triple by using triple-index scaling, calculated value sharing, and symbol degradation algorithms, obtains the fiber optic nonlinear damage value by equalizing the simplified triple through the CLDNN network, and also applies a part of the parameters trained in the source domain to the target domain through transfer learning, reducing the computational complexity of the algorithm. This solution can not only obtain high system transmission performance with low computational complexity, but also make the compensation solution have high universality. Brief Description of the Drawings

[0022] Figure 1 It is a single-channel 30 GBaud DP-16QAM coherent optical transmission system diagram.

[0023] Figure 2 It is a symbol degradation block diagram.

[0024] Figure 3 It is a fiber optic nonlinear equalization structure diagram based on the TL-CLDNN algorithm. Detailed Embodiment

[0025] The following further describes the present invention in detail with reference to the drawings and specific embodiments. The technical solution adopted by the present invention is:

[0026] First, obtain the transmission data of the coherent optical fiber communication system: construct a DP-16QAM coherent optical fiber communication system, and the transmission system is as Figure 1 shown. At the sending end, the bit sequence has a length of 2 16Pseudo-random bit sequence generation. First, a 30 GBaud single-channel DP-16QAM symbol with root-raised cosine 0.01 Nyquist pulse shaping is generated using a 60 Gsample / s arbitrary waveform generator; then, a polarization beam splitter splits the 1550 nm optical carrier generated by an external cavity laser into two polarized lights, and the corresponding IQ modulators are used to modulate and generate modulation signals respectively. Finally, a polarization beam combiner combines the two modulation signals into one optical signal, which is then power-amplified by an erbium-doped fiber amplifier and transmitted into the fiber link. On the fiber link side, the link is implemented through a fiber loop, which consists of 100 km of standard single-mode fiber, an erbium-doped fiber amplifier, a band-pass filter, and a loop controller. At the receiving end, first, the out-of-band noise is filtered out by a band-pass filter, and then IQ imbalance compensation, dispersion compensation, polarization demultiplexing, frequency offset estimation, and carrier phase recovery are performed.

[0027] Then, through the first-order perturbation solution of the nonlinear Schrödinger equation, the simplified triple is calculated.

[0028] Finally, the fiber nonlinearity is obtained by equalizing the simplified triple through the TL-CLDNN network, thereby compensating the coherent optical fiber communication system.

[0029] The specific steps of the fiber nonlinear equalization method based on the TL-CLDNN algorithm of the present invention are as follows:

[0030] Step 1: Calculate the product of the first two symbols in each triple through the triple derived from the first-order perturbation solution of the nonlinear Schrödinger equation, and the result is shared by multiple consecutive symbols.

[0031] The specific process of calculating and sharing the product of the first two symbols in the triple is as follows:

[0032] Based on the Manakov equation, the transfer function under double polarization of a single-mode fiber can be expressed by the NLSE as:

[0033]

[0034] where u x / y (z, t) represents the optical field of the signal in the x or y polarization direction; α represents the fiber attenuation coefficient; β2 represents the group velocity dispersion; γ represents the fiber nonlinear coefficient. Based on the first-order perturbation theory, the in-channel fiber nonlinear distortion of the double-polarization symbol is approximately expressed as:

[0035]

[0036] where △u k,x and △u k，y are the fiber nonlinear perturbations in the x polarization and y polarization at the k-th moment respectively; P0 is the transmitted power; X k+m and Yk+m They represent the transmission symbols of x polarization and y polarization at the (k+m)th moment respectively; and They represent the conjugate of the transmission symbol of x polarization and y polarization at the (k+m+n)th moment respectively; m and n are symbol indices; C m，n represents the nonlinear coefficient of the optical fiber; at this time, the calculation formula of the triplet is:

[0037]

[0038] Where, T k，x and T k，y They represent the triplets on x polarization and y polarization at time k respectively. Then, the selection criteria of index coefficients in the triplets are established, and the criteria are as follows:

[0039]

[0040] In the formula, the value of L depends on the size of the channel storage capacity; R is used to balance the performance and computational complexity of fiber nonlinear equalization. After determining the triple index, calculate X at time k k+n X k+m , Y k+n Y k+m , Y k+n X k+m and X k+n Y k+m The calculated value can be shared with other moments.

[0041] Step 2: Degenerate the remaining symbols in the triple and replace the second multiplication with a simple logical operation.

[0042] Based on the semi-degenerate theory, and The symbol degenerates to a QPSK-like symbol and origin, and the symbol degradation diagram is shown in Figure 2 Therefore, logical operations can be used instead of X k+n X k+m , Y k+n Y k+m , Y k+n X k+m , X k+n Y k+m and The complex multiplication operation between them. At this time, the intra-channel fiber nonlinear distortion of the dual polarization symbols can be approximately expressed as:

[0043]

[0044] In the formula, Denote the perturbation coefficient after N-level quantization; Log(·) represents the logical operation of symbol degradation. At this time, the calculation formula of the simplified triple is as follows:

[0045]

[0046] Step 3: Cascade the CNN, LSTM, and DNN networks to construct a unified CLDNN network, and equalize the simplified triple to obtain the optical fiber nonlinear damage, so as to compensate the coherent optical fiber communication system.

[0047] As Figure 3 shown, input the feature triple at the current moment and the corresponding label into the CLDNN network. After equalization processing, the estimated nonlinear damage is obtained and removed from the received symbol to obtain the compensated symbol. Finally, calculate the bit error rate and Q value of the transmission system.

[0048] Step 4: Adopt transfer learning based on multi-source domain, apply a part of the parameters trained in the source domain to the target domain, and quickly compensate the target coherent optical fiber communication system.

Claims

1. A fiber optic nonlinear equalization method based on the TL-CLDNN algorithm, characterized in that, Including the following steps: Step 1: Construct a DP-16QAM coherent optical fiber communication system; Step 2: Selection and calculation of triples; Step 3: Construct a TL-CLDNN network; Step 4: Optical fiber nonlinear equalization based on the TL-CLDNN algorithm.

2. The fiber optic nonlinear equalization method based on the TL-CLDNN algorithm according to claim 1, characterized in that The DP-16QAM coherent optical fiber communication system in the said Step 1 includes: At the transmitting end, the bit sequence is generated from a pseudo-random bit sequence with a length of 2 16 . First, a 30 GBaud DP-16QAM symbol with root-raised cosine 0.01 Nyquist pulse shaping is generated using a 60 Gsample / s arbitrary waveform generator; then, a polarization beam splitter splits the 1550 nm optical carrier generated by an external cavity laser into two polarized lights, and the corresponding IQ modulators are used to modulate and generate modulation signals respectively. Finally, a polarization beam combiner combines the two modulation signals into one optical signal, which is then power-amplified by an erbium-doped fiber amplifier and transmitted into the fiber link. On the fiber link side, the link is implemented through a fiber loop, which consists of 100 km of standard single-mode fiber, an erbium-doped fiber amplifier, a band-pass filter, and a loop controller. At the receiving end, first, the out-of-band noise is filtered out by a band-pass filter, and then IQ imbalance compensation, dispersion compensation, polarization demultiplexing, frequency offset estimation, and carrier phase recovery are performed.

3. A fiber optic nonlinear equalization method based on the TL-CLDNN algorithm according to claim 1, characterized in that The selection and calculation process of the triples in the said Step 2 includes: Based on the Manakov equation, the transfer function under the double polarization of a single-mode optical fiber can be expressed by the NLSE as: where \(u\) x / y (z, t) represents the optical field of the signal in the x or y polarization direction; α represents the fiber attenuation coefficient; β2 represents the group velocity dispersion; γ represents the fiber nonlinear coefficient. Based on the first-order perturbation theory, the fiber nonlinear distortion in the channel with dual-polarization symbols is approximately expressed as: where Δu k，x and Δu k，y are the fiber nonlinear perturbations on the x - polarization and y - polarization at time k, respectively; P0 is the transmitted power; X k+m and Y k+m represent the transmission symbols on the x - polarization and y - polarization at time (k + m), respectively; and represent the conjugates of the transmission symbols on the x - polarization and y - polarization at time (k + m + n), respectively; m and n are symbol indices; C m,n represents the nonlinear coefficient of the optical fiber. At this time, the calculation formula of the triple is: where T k，x and T k，y represent the triplets of the x-polarization and y-polarization at the k-th moment, respectively. Then, a selection criterion for establishing the triplet index is as follows: In the formula, the value of L depends on the size of the channel storage capacity; R is used to balance the performance and computational complexity of fiber nonlinear equalization. After determining the triple index, calculate X at time k k+n X k+m , Y k+n Y k+m , Y k+n X k+m and X k+n Y k+m The value of , and the calculated value is shared with other moments. Then, the symbol and Degenerate into QPSK signal or origin, replace X by simple logic operation k+n X k+m , Y k+n Y k+m , X k+n Y k+m and Y k+n X k+m and and At this time, the simplified calculation formula of the triple is: In the formula, Log(·) represents the logical operation of symbol degradation.

4. A fiber optic nonlinear equalization method based on the TL-CLDNN algorithm according to claim 1, characterized in that, The said Step 3 specifically includes: Step 3.1: Cascade the CNN, LSTM, and DNN networks to construct a unified CLDNN network; Step 3.2: Adopt transfer learning based on multi-source domains, and apply a part of the parameters trained in the source domain to the target domain.

5. A fiber optic nonlinear equalization method based on the TL-CLDNN algorithm according to claim 1, characterized in that, The TL-CLDNN network equalizes the simplified triples to obtain the estimated optical fiber nonlinear damage value, and uses this value to compensate for the nonlinear damage of the coherent optical fiber communication system.

Images