Optical fiber communication system nonlinear compensation module based on photon chip

Abstract

The invention provides an optical fiber communication system nonlinear compensation module based on a photon chip, and belongs to the field of optical fiber communication. Comprising a photon chip, an FPGA module, an ADC / DAC and a light source, and the photon chip, the FPGA module, the ADC / DAC and the light source are utilized to build a full-connection neural network used for calculating nonlinear disturbance signals in the optical fiber communication system and compensating optical signals at a receiving end so as to improve the transmission performance of the optical fiber communication system. According to the invention, the cascaded Mach-Zehnder Interferometer (MZI) array and the FPGA electronic control unit are utilized to construct a photon neural network computing system together, so that neural network operation can be executed at high speed and low power consumption. The non-linear damage compensation module is used for compensating non-linear damage SPM, IXPM and IFWM generated in the single-mode fiber transmission process, and can be used as a non-linear damage compensation module of a digital signal processing technology in an optical fiber communication system so as to improve the transmission performance of the optical fiber communication system.

Description

Technical Field

[0001] The present invention relates to the fields of optical fiber communication and photonic artificial intelligence, and in particular to a nonlinear compensation module of an optical fiber communication system based on a photonic chip. Background Art

[0002] The rapid development of technologies such as big data, cloud computing, and artificial intelligence has driven the explosive growth of the capacity demand of optical communication systems, while nonlinear damage is the main factor limiting the further improvement of the capacity of optical communication systems. Digital signal processing technology provides a flexible and efficient way to achieve compensation for optical signals in the electrical domain, which can greatly improve the transmission capacity of optical fiber communication systems.

[0003] In high-speed fiber-optic communication systems, the most important nonlinear effects are cross-phase modulation (IXPM) and four-wave mixing (IFWM) within the channel. IXPM means that in a dual-polarization system, an optical signal in one polarization direction can affect an optical signal in another polarization direction through the nonlinear optical Kerr effect, thereby causing channel crosstalk problems. IFWM can cause fluctuations in pulse amplitude, which may cause parasitic optical pulses to be generated at the "0" code, which originally had no optical power, causing the pulse amplitude at the "1" code to change, resulting in power exchange between pulses. Neural networks can achieve efficient nonlinear damage compensation. In the training phase, the receiver learns from the training data to generate a black box model of the transmission link. In the execution phase, the nonlinear damage of the system is calculated based on the model and removed from the data, thereby achieving longer transmission distances and higher transmission capacity.

[0004] Electronic neural network hardware usually uses the von Neumann architecture, which separates storage and computing. When executing operations, data must be repeatedly read and written between the central processing unit and the memory, which limits the computing speed and energy consumption of a single operation. Photonic neural networks use optical signals as a carrier for information transmission, and can directly perform high-speed data processing in the analog domain. The energy consumption of optical devices is lower than that of electronic devices. Therefore, the photonic chip architecture designed to achieve specific functions can achieve high-speed and low-power neural network operations.

[0005] The neural network based on perturbation theory regards the nonlinear damage perturbation term as a shadow pulse generated by adjacent pulses. The perturbation term can be calculated based on the interaction between adjacent pulses and compensated at the receiving end. Using the photonic neural network chip as hardware support, a two-layer fully connected neural network is built, and a compensation system based on perturbation theory is designed to learn from the received optical signal to compensate for nonlinear damage. Summary of the invention

[0006] The present invention provides a nonlinear compensation module for an optical fiber communication system based on a photonic chip. The photonic chip based on an MZI cascade network performs the main data processing process, and together with the optical computing control module composed of an FPGA, an ADC, and a DAC, a neural network is implemented, and nonlinear damage in the transmission process is compensated based on the received optical signal. The advantages of the large bandwidth, high speed, and low energy consumption of the photonic chip are utilized to achieve high-speed digital signal processing.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A nonlinear compensation module for an optical fiber communication system based on a photonic chip, comprising:

[0009] The photonic chip is used to perform the linear operation part of the neural network. It realizes any weight matrix by changing the voltage signal applied to the phase shifter, completes the weighted multiplication process of the input signal and the weight matrix, and the detector array integrated on the chip realizes the summation process.

[0010] The FPGA module is responsible for controlling the timing of the optical signal to perform linear operations on the photonic chip and nonlinear operations on the FPGA. It can control the electrical signals applied to the photonic chip, implement the nonlinear activation function in the neural network, and send the processed data back to the host computer to output the calculation results.

[0011] ADC / DAC, connected to the interface of the FPGA and the photonic chip, providing four-channel AD / DA conversion for analog-to-digital conversion and digital-to-analog conversion between the optical computing module and the FPGA;

[0012] Light source, providing optical signal input for the system;

[0013] The photonic neural network uses the above-mentioned photonic chip, FPGA module, ADC / DAC and light source to build a fully connected neural network, which is used to calculate the nonlinear disturbance signal in the optical fiber communication system and compensate the optical signal at the receiving end to improve the transmission performance of the optical fiber communication system.

[0014] Furthermore, the photonic chip is composed of a cascaded MZI array, and the optical signal is modulated into an ideal output signal through the cascaded MZI array to complete the process of multiplying the input vector and the weight matrix in the neural network, and the large bandwidth characteristics of the photonic chip are used to achieve high-speed neural network operations. Among them, each of the inner and outer arms of the MZI contains a thermo-optical phase shifter, which can change the voltage value applied thereto, and utilizes the characteristic that the refractive index of silicon materials changes with temperature. By introducing a microheater in the device, the external electric field energy is converted into thermal energy, thereby changing the refractive index of the material and achieving modulation of the light field.

[0015] Further, the ADC / DAC includes:

[0016] High-speed ADC, which realizes four-channel AD conversion, is used to convert the analog signal output by the photonic chip into a digital signal for sending to the FPGA to perform the next calculation process;

[0017] High-speed DAC realizes four-channel DA conversion, which is used to convert the digital electrical signals output by FPGA into analog electrical signals to control the photonic chip.

[0018] Furthermore, the photonic neural network is composed of two hidden layers, the input data is the three products of the symbols collected after optical fiber transmission, and the output result is the nonlinear damage term of the link, which includes self-phase modulation SPM in the optical fiber channel, cross-phase modulation XPM in the channel, and four-wave mixing IFWM in the channel.

[0019] Compared with the existing nonlinear compensation module, the embodiments of the present invention have the following beneficial effects: the present invention directly realizes the nonlinear damage calculation of the received symbols in the optical domain by using FPGA, photonic chip, etc., fully utilizing the advantages of optical computing and realizing the true meaning of "propagation is computing". In addition, the optical device is easy to connect with the optical fiber communication system to realize the integration of data generation, transmission and processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 System diagram of nonlinear compensation module of optical fiber communication system based on photonic chip;

[0021] Figure 2 MZI structure diagram;

[0022] Figure 3 Photonic chip design diagram;

[0023] Figure 4 Fiber optic communication system nonlinear compensation module processing flow chart; DETAILED DESCRIPTION

[0024] The technical scheme and hardware structure in the embodiments of the present invention will be fully described below in conjunction with the drawings in the embodiments of the present invention. Unless otherwise specifically stated, the relative arrangement of the components and steps described in these embodiments does not limit the scope of the present invention. In order to make the objects, technical schemes and advantages of the present invention more obvious, the exemplary embodiments according to the present invention will be described in detail below with reference to the drawings. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments of the present invention, and it should be understood that the present invention is not limited to the exemplary embodiments described herein. Based on the embodiments of the present invention described in the present invention, all other embodiments obtained by those skilled in the art without paying creative work should fall within the protection scope of the present invention.

[0025] Compared with electronic nonlinear compensation modules, photonic computing chips have greater bandwidth and lower energy consumption. After training, the electrical signals on the optical chip no longer need to be adjusted, and the number of operations that can be performed per unit of energy consumption is 4 to 5 orders of magnitude higher than that of typical electronic devices such as CPUs and GPUs. Therefore, using photonic computing to replace traditional electronic computing modules to perform the nonlinear compensation process of optical fiber communication systems can alleviate the computing power and power consumption pressure of digital signal processing technology in the current optical fiber communication field.

[0026] Reference Figure 1 The optical computing system provided by the present invention includes a photonic chip composed of an MZI cascade array and a detector, a light source, a polarization controller, an FPGA, an ADC / DAC, and a temperature control unit. The data sent to the neural network is a signal modulated by the light source. The polarization controller controls the strength of the light signal entering the chip. The photonic chip is responsible for performing the neural network linear matrix multiplication operation. The detector is an on-chip photodetector based on Ge material. The FPGA performs electrical control and sends the processed data back to the host computer. Specifically, it includes the following implementation methods:

[0027] Photonic chip design: Reconfigurable MZI can be used to build a 2×2 optical signal linear processor, which can realize the N-dimensional special unitary group transformation matrix SU(N) by N(N-1) / 2 MZIs in N channels. Each MZI in this structure consists of two 3dB directional couplers. The thermo-optical phase shifter on the inner arm of the MZI is used to control the power ratio of the output end, and the phase shifter on the outer arm can adjust the relative phase of the output end. To realize the weight matrix in the neural network using the MZI cascade network, it is necessary to first decompose any weight matrix into the result of multiplying the unitary matrix with the diagonal matrix and the unitary matrix according to the singular value diversity theory. Then, N(N-1) / 2 MZI cascades are used to realize the two unitary matrices after decomposition, and then the MZI with a single input / output port is used as an optical attenuator to realize the diagonal matrix. They are connected in sequence on the optical architecture. The process of the optical signal passing through the cascade architecture completely corresponds to the result of multiplying the input signal of the neural network with the weight matrix. An MZI cascade network based on silicon waveguide is designed, and the corresponding photonic chip is obtained through SOI process.

[0028] FPGA-based optical computing control unit: Optical computing control is the core module that supports the normal operation of the photonic computing system. It is responsible for controlling the information transmission, timing interruption, data processing and other processes within the chip. It also undertakes the data calculation and voltage signal control functions in the photonic neural network training phase. For the received optical signal, the FPGA needs to encode it as the input data of the neural network, send it to the network to perform weight matrix operations, and perform the nonlinear activation mapping process of the neural network after the linear operation is completed to enhance the learning ability of the network model. After the neural network operation process is completed, the FPGA needs to send the optical signal back to the host computer through the ADC and output the network calculation results.

[0029] Design of neural network model based on perturbation theory: For nonlinear Kerr damage in optical fiber communication systems, perturbation theory transmits the current signal pulse and the adjacent signal pulse into the neural network model, and expresses the nonlinear damage as the nonlinear phase disturbance of the pulse envelope at the current position caused by pulses at three different times. Through the neural network, the nonlinear phase disturbance to each symbol can be analyzed. Compared with the traditional perturbation scheme, the advantage of the neural network perturbation compensation scheme is that as the transmission distance increases, the computational complexity of the traditional perturbation scheme will increase rapidly, resulting in an increase in computing time and energy consumption. However, after the neural network training is completed, the number of calculations does not increase with the increase in transmission distance, which is more conducive to the implementation of high-speed, long-distance optical fiber communication system nonlinear compensation.

[0030] Computing system integration: After the above functional parts are designed, considering the integration of the computing module, the light source, FPGA, ADC, DAC, photonic chip, temperature control and other modules need to be assembled into a chassis to build a complete computing system. Among them, the light source is responsible for modulating the optical signal into symbols sent to the neural network, the photonic chip performs the linear operation part of the neural network, and the ADC and DAC are responsible for the conversion process between digital signals and analog signals. Among them, the high-speed ADC realizes four-channel AD conversion, which is used to convert the analog signal output by the photonic chip into a digital signal for sending to the FPGA to perform the next operation process. The high-speed DAC realizes four-channel DA conversion, which is used to convert the digital electrical signal output by the FPGA into an analog electrical signal to control the photonic chip. The FPGA controls the timing and performs nonlinear operations of the neural network. The optical signal is transmitted in the linear and nonlinear modules in turn, and finally forms a two-layer fully connected neural network to calculate the nonlinear perturbation damage caused by the link. The signal is compensated in the host computer, and finally the bit error rate is reduced and the performance of the communication system is improved.

[0031] In the system of the present invention, at the receiving end, the incident signal is first coherently mixed with the local oscillator optical signal to achieve coherent detection of in-phase and orthogonal signals. Then, digital signal processing is performed after photoelectric conversion and digital-to-analog conversion: linear damage compensation of the signal is completed after RRC filtering, and then polarization demultiplexing is achieved according to the CMA algorithm, and carrier phase recovery is achieved using the VV algorithm. Nonlinear damage is calculated based on the neural network and compensated before symbol judgment, and finally the compensation and decoding process of the optical fiber communication system signal is completed.

[0032] Based on the above implementation modes, the photonic chip in the present invention is mainly composed of a MZI cascade array. Figure 2 The MZI unit structure is shown. Figure 3 It is the arrangement diagram of the MZI cascade array in the photonic chip. The optical signal is modulated into an ideal output signal through the cascaded MZI array, which is equivalent to completing the process of multiplying the input vector and the weight matrix in the neural network. The large bandwidth characteristics of the photonic chip are used to realize high-speed neural network operations. The optical principle of the MZI unit is: two light waves are input into the MZI from the upper and lower arms, and after passing through the first 3dB directional coupler, they become two signals with the same light intensity. One of the light waves passes through the thermo-optical phase shifter θ. At this time, the thermo-optical phase shifter will change the phase of the light wave according to the electrical signal input to the phase shifter, and the phase of the other light wave remains unchanged. After that, the two light waves will be transmitted to the second directional coupler, and then the relative phase will be changed by the thermo-optical phase shifter φ. In the present invention, each of the inner and outer arms of the MZI includes a thermo-optical phase shifter, which can change the voltage value applied thereto. By utilizing the characteristic that the refractive index of silicon material changes with temperature, a microheater is introduced into the device to convert the external electric field energy into thermal energy, thereby changing the refractive index of the material and realizing the modulation of the light field.

[0033] After the optical signal passes through the MZI, the intensity and phase of the output optical signal change, which can be regarded as the digital information carried by the light wave completing a multiplication operation. After passing through a cascade array composed of multiple MZIs, matrix multiplication operations can be realized.

[0034] In optics, the product of matrices is equivalent to the input signal passing through the experimental device in the order of the multiplication formula. Since the transmission matrix of the MZI in the N-dimensional Hilbert space only affects one of the two-dimensional subspaces, the remaining (N-2)-dimensional subspaces remain unchanged. From an optical point of view, each time a MZI passes through, only the two signals connected to the two ports of the MZI change, and the remaining signals remain unchanged. When many MZI units are cascaded into a network, continuous two-dimensional matrix transformations can be performed on the two-dimensional subspace of the N-dimensional Hilbert space. That is, by cascading multiple MZIs into a multi-channel network, any SU(N) optical transformation can be realized in the N-dimensional space. Therefore, finding an optical architecture that realizes an arbitrary weight matrix is ​​equivalent to mathematically decomposing the weight matrix into the product of unitary matrix-diagonal matrix-unitary matrix, and building a corresponding experimental device in the optical domain.

[0035] After the architecture of the photonic chip is designed, it can be outsourced using the 130nm silicon photonics process, which can achieve a 130nm line width and support double-layer metal copper interconnection.

[0036] In addition to controlling the timing and the electrical signals applied to the photonic chip, the FPGA also performs the nonlinear mapping process of the neural network and implements the ReLU and Softmax functions.

[0037] In the optical fiber communication system, the position of this module in the digital signal processing flow is as follows: Figure 4 As shown, it undertakes the nonlinear compensation function. This module processes the optical signal at the receiving end and can learn from the symbols received by the receiver to build a reasonable model without relying on prior knowledge of the link.

[0038] According to perturbation theory, the nonlinear term is caused by the joint influence of three pulses at different times on the current signal, so the input data of the photonic neural network is the triple product "triplet" obtained by multiplying three symbols.

[0039] The specific details of the neural network are as follows:

[0040] The neural network consists of two hidden layers. The input data is the three products of the symbols collected after optical fiber transmission. The output result is the nonlinear damage term of the link, which includes self-phase modulation SPM in the optical fiber channel, cross-phase modulation IXPM in the channel, and four-wave mixing IFWM in the channel.

[0041] For the collected symbol data, 60% is taken as the training set, and the remaining 20% ​​is taken as the cross-validation set and the test set;

[0042] The processing layer of the neural network consists of two layers, with 8 and 2 neurons respectively;

[0043] A dropout layer is set to prevent overfitting, with a parameter of 0.5;

[0044] The batch size is 10;

[0045] The optimization algorithm is Adam optimizer, and the learning rate is 0.001;

[0046] The loss function is the mean square error between the compensated symbol and the standard symbol at the transmitter;

[0047] The output structure of the neural network is a nonlinear damage size.

[0048] This photonic neural network only realizes nonlinear damage compensation in the digital signal processing process, and still needs to be combined with other modules in the electrical domain to realize the complete signal processing process.

[0049] During the entire process, the photonic chip performed the main nonlinear compensation process and realized some digital signal processing functions in the fiber-optic communication system, with high speed, low consumption and flexible training.

[0050] The implementation of the present invention can not only be used for nonlinear damage compensation in optical fiber communication systems, but also can be applied to other digital signal processing tasks based on neural networks, such as dispersion equalization, symbol decision, etc., and has broad application prospects. The above description of the specific implementation methods is intended to describe and illustrate the technical solutions of the present invention. The above specific implementation methods are only illustrative and not restrictive. Without departing from the scope of protection of the present invention and the claims, ordinary technicians in this field can also make many forms of specific changes under the enlightenment of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A nonlinear compensation module for optical fiber communication system based on photonic chip, characterized in that include: The photonic chip is used to perform the linear operation part of the neural network. It realizes any weight matrix by changing the voltage signal applied to the phase shifter, completes the weighted multiplication process of the input signal and the weight matrix, and the detector array integrated on the chip realizes the summation process. The FPGA module is responsible for controlling the timing of the optical signal to perform linear operations on the photonic chip and nonlinear operations on the FPGA. It can control the electrical signals applied to the photonic chip, implement the nonlinear activation function in the neural network, and send the processed data back to the host computer to output the calculation results. ADC / DAC, connected to the interface of the FPGA and the photonic chip, providing four-channel AD / DA conversion for analog-to-digital conversion and digital-to-analog conversion between the optical computing module and the FPGA; Light source, providing optical signal input for the system; The photonic neural network uses the above-mentioned photonic chip, FPGA module, ADC / DAC and light source to build a fully connected neural network, which is used to calculate the nonlinear disturbance signal in the optical fiber communication system and compensate the optical signal at the receiving end to improve the transmission performance of the optical fiber communication system.

2. The optical fiber communication system nonlinear compensation module according to claim 1, characterized in that: The photonic chip is composed of a cascaded MZI array. The optical signal is modulated into an ideal output signal through the cascaded MZI array, completing the process of multiplying the input vector and the weight matrix in the neural network. The large bandwidth characteristics of the photonic chip are used to achieve high-speed neural network operations.

3. The optical fiber communication system nonlinear compensation module according to claim 1, characterized in that: The ADC / DAC comprises: High-speed ADC, which realizes four-channel AD conversion, is used to convert the analog signal output by the photonic chip into a digital signal for sending to the FPGA to perform the next calculation process; High-speed DAC realizes four-channel DA conversion, which is used to convert the digital electrical signals output by FPGA into analog electrical signals to control the photonic chip.

4. The optical fiber communication system nonlinear compensation module according to claim 2, characterized in that: The inner and outer arms of the MZI each contain a thermo-optic phase shifter that can change the voltage applied to it. By utilizing the property that the refractive index of silicon material changes with temperature and introducing a microheater in the device, the external electric field energy is converted into thermal energy, thereby changing the refractive index of the material and achieving modulation of the light field.

5. The optical fiber communication system nonlinear compensation module according to claim 1, characterized in that: The photonic neural network consists of two hidden layers. The input data is the three products of the symbols collected after optical fiber transmission. The output result is the nonlinear damage term of the link, which includes self-phase modulation SPM in the optical fiber channel, cross-phase modulation IXPM in the channel, and four-wave mixing IFWM in the channel.

Images