Global optimization method for optical communication systems based on neural network and genetic algorithm

Abstract

The present invention proposes a global optimization method for optical communication systems that combines a neural network with a genetic algorithm to accelerate the optimal configuration of system parameters. On the one hand, the present invention models the system through a pre-finite impulse response (FIR) filter and a clipping neural network to obtain an offline model with configurable FIR tap coefficients, clipping coefficients, and DAC peak-to-peak values, so that the parameter optimization process can be completed without affecting the actual system operation. On the other hand, by combining the characteristics of the genetic algorithm (GA) and the back propagation (BP), the two are used alternately in different optimization periods to perform multi-dimensional parameter optimization to accelerate the global optimal search for parameters. In this way, in optical communication systems with severe transceiver band limitations, the optimal configuration of system parameters is achieved, and the system's ability to resist high-frequency fading is improved.

Description

Technical Field

[0001] The present invention relates to optical fiber communication technology, in particular to a global optimization technology for a high baud rate optical communication system. Background Art

[0002] In recent years, with the rapid development of technologies such as cloud computing, data centers, and ultra-high-definition video, global network data traffic has experienced explosive growth. This trend has driven the development of fiber-optic communication systems towards greater capacity and higher transmission rates. In particular, as baud rates exceed 100GBaud, the bandwidth limitations of transceiver electronics become increasingly severe.

[0003] In order to overcome the band limitation of the transceiver and improve the system performance, on the one hand, digital pre-emphasis DPE technology can be used at the transmitter to enhance the high-frequency components of the signal. However, DPE will increase the peak-to-average power ratio PAPR of the signal, but in a peak-limited system, it will reduce its average power, which ultimately reduces the signal-to-noise ratio SNR of the system. On the other hand, at the receiver, a feedforward equalizer FFE can be used to compensate for the high-frequency components of the signal, but at the same time it will cause the high-frequency noise to be enhanced, affecting the system performance. Therefore, it is necessary to reasonably allocate the degree of compensation of each transceiver, that is, to optimize the tap coefficient of the transmitter DPE. Considering that the use of DPE or FFE alone has its limitations, partial DPE has been proposed. Among the partial DPE solutions, one is to use a two-tap filter 1+αz -1 , by scanning the tap coefficient α to determine the value when the system performance is optimal. Another solution is partial channel compensation, that is, setting the pre-emphasis frequency response to |H -β , a Fourier transform yields the time-domain tap coefficients, where H represents the channel frequency response and β represents the compensation level. The optimal value requires scanning. These are all single-variable optimization schemes with limited optimization capabilities. Furthermore, due to the high PAPR of the signal after DPE, clipping is required. The clipped digital signal passes through the digital-to-analog converter (DAC), and the peak-to-peak value (Vpp) of the DAC output signal must be controlled. Excessively high Vpp can cause nonlinearity in subsequent devices, while too low a Vpp can result in low signal power, impacting the signal-to-noise ratio. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for optimizing the configuration of multi-dimensional system parameters such as FIR tap coefficients, clipping coefficients, and DAC peak-to-peak values ​​by performing high-precision and rapid modeling of an offline system, in order to address the problem that in a high-baud-rate optical communication system, the limited transceiver bandwidth and severe high-frequency fading make it difficult to optimize the configuration of the system's multi-dimensional parameters, thereby affecting the system's signal-to-noise ratio.

[0005] The technical solution adopted by the present invention to solve the above technical problems is a global optimization method for optical communication systems using a neural network combined with a genetic algorithm, comprising the following steps:

[0006] Steps to establish an offline optical communication system model:

[0007] Determining a parameter module to be configured in an optical communication system; using multi-dimensional parameters to represent the parameter to be configured, where N parameters to be configured have N dimensions; and training a neural network to simulate a parameter module not to be configured in the optical communication system;

[0008] The parameter module to be configured is placed before the trained neural network to form an offline optical communication system model;

[0009] Multi-dimensional parameter alternating update steps:

[0010] A genetic algorithm (GA) is used to update the multidimensional parameters in one round, wherein the received BP-optimized multidimensional parameters are used as the first individual in the initial population; the genetic algorithm (GA) outputs the multidimensional parameters corresponding to the optimal individual in the last population as the GA-optimized multidimensional parameters to the back propagation algorithm (BP); the GA-optimized multidimensional parameters are used as the BP initial parameter values, and BP is used to update the multidimensional parameters in one round, outputting the BP-optimized multidimensional parameters to complete a multidimensional parameter alternation update;

[0011] Determine whether the currently updated multidimensional parameters meet the alternating update end condition. If so, output the latest multidimensional parameters as the multidimensional parameters for global optimization. Otherwise, repeat the alternating update of the multidimensional parameters.

[0012] Actual configuration steps: Use the globally optimized multi-dimensional parameters to configure a real optical communication system to complete the global optimization of the optical communication system.

[0013] Specifically, the optical communication system is an intensity modulation direct detection (IMDD) system; the parameters that need to be configured include FIR tap coefficients, clipping coefficients, and DAC peak-to-peak value Vpp; the parameter modules to be configured include an FIR filter, a clipping model, and a digital-to-analog converter DAC; the DAC peak-to-peak value Vpp is input into the neural network as a configurable parameter.

[0014] In order to flexibly configure and optimize parameters, the present invention, on the one hand, models the system through a pre-FIR filter (DPE module) and a Clipping neural network to obtain an offline model with configurable FIR tap coefficients, Clipping coefficients, and DAC peak-to-peak values, so that the parameter optimization process can be completed without affecting the actual system operation. On the other hand, due to the timeliness of communication needs, the system has high requirements on the convergence speed of the algorithm. Genetic algorithm GA is a method of searching for the global optimal solution by simulating the natural evolution process, but the algorithm converges slowly. The back propagation algorithm BP based on neural network guides the parameter optimization direction by calculating the gradient. The algorithm converges quickly, but it is easy to fall into local optimality. The present invention combines the advantages of the above two algorithms to achieve fast global optimal search capabilities, thereby completing the fast global optimization of various parameters in a severely band-limited optical communication system and improving system performance.

[0015] The beneficial effects of the present invention are as follows: first, an offline model of configurable system parameters is established, and then the characteristics of the GA and BP algorithms are combined and used alternately to accelerate the global optimal search for parameters, making full use of the respective advantages of the two algorithms and avoiding their disadvantages, thereby achieving high-complexity, multi-dimensional parameter optimization configuration of the entire system, improving the performance of optical communication systems with severe transceiver band limitations, and meeting the timely global optimization requirements of parameters in high-speed optical communication systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is the GA and BP alternating optimization flow chart of the present invention.

[0017] Figure 2 Schematic diagram of the system modeling and parameter global optimization scheme of the present invention. DETAILED DESCRIPTION

[0018] The global optimization method for optical communication systems using a neural network combined with a genetic algorithm first completes offline optical communication modeling using a neural network prepended with the module containing the parameters to be configured, resulting in an offline optical communication model of the configurable system. Based on this offline optical communication model, the parameters to be configured are then optimized in a step-by-step alternating manner using a combination of the GA and BP algorithms. This approach achieves global search and local rapid convergence, ultimately achieving optimal optimization of the parameters to be configured. There can be multiple parameters to be configured, each corresponding to a dimension. N parameters to be configured have N dimensions, which is called a multidimensional parameter set. N is a positive integer greater than or equal to 1.

[0019] The specific steps of combining GA and BP algorithms are as follows: Figure 1 As shown:

[0020] Genetic algorithm GA: Generates the initial population, and the first individual in the population is replaced by the multidimensional parameters output by the BP algorithm. If the GA algorithm has not undergone BP optimization in the first round, it will not be replaced. Each individual in the population is a set of parameters to be optimized, that is, a multidimensional parameter, whose value is randomly generated within the allowable range. Configure each set of parameters in the population to the system model in turn, and calculate the MSE of the model input and output signals under each set of parameters. Select the three individuals with the highest fitness in the population, cross them with each other, and then mutate them. The new individuals obtained form a new population with the original three individuals. Then repeat the above operations of calculating fitness, selecting, crossing, mutating, and generating a new population five times, and output the multidimensional parameters corresponding to the optimal individual in the population for the last time;

[0021] Back Propagation Algorithm (BP): The multidimensional parameters corresponding to the optimal individual output by the GA algorithm are used as the initial parameter values ​​and configured into the system model. The MSE of the model input and output signals under the current multidimensional parameter configuration is calculated. The gradient of the MSE function with respect to the multidimensional parameters is obtained through the back propagation method of the neural network. The multidimensional parameters are then updated using the gradient descent method. The above operation is repeated until the MSE converges and the final multidimensional parameters are output.

[0022] Repeat the GA and BP steps alternately for multiple times until the MSE converges and the performance is stable, that is, the optimal parameter value is output. Figure 1 .

[0023] Example

[0024] like Figure 2 As shown in the figure, for the intensity modulation direct detection IMDD system in actual optical communication, it consists of a transmitting end consisting of an FIR filter, a clipping module, a digital-to-analog converter DAC, an electrical amplifier, a laser and a Mach-Zehnder modulator MZM, an optical fiber, and a receiving end consisting of a photodetector PD, an analog-to-digital converter ADC and a feedforward equalizer FFE.

[0025] At the transmitter, the pulse amplitude modulated (PAM) signal undergoes FIR filtering and clipping processing, then passes through a DAC to generate an analog electrical signal, which is then amplified by an electrical amplifier. The amplifier's electrical signal, along with the laser output, is then input into the MZM for modulation to produce an optically modulated signal. The optical signal emitted by the MZM then travels through optical fiber to the receiver. At the receiver, the optical signal sequentially passes through the PD, ADC, and FFE to recover the original PAM signal.

[0026] The optimization of the IMDD system in the embodiment is to configure the optimal FIR filter tap coefficients at the transmitter, the Clipping coefficients of the Clipping module and the DAC peak-to-peak value Vpp input to the DAC. Before optimizing the parameters, an IMDD system offline model is first established, such as Figure 2 As shown in Figure 2, only the FIR filter and Clipping module are retained to facilitate the configuration of the parameters to be optimized. The remaining modules in the IMDD system are all modeled using a deep neural network. FIR is a feedforward N-tap filter with tap coefficients [w0, w1, ..., w N-1 ] is a configurable parameter. The other two configurable parameters are Clipping coefficient and DAC peak-to-peak value Vpp.

[0027] On the established IMDD system offline model, a parameter optimization method combining GA and BP is used to achieve fast global optimization of multi-dimensional parameters. The multi-dimensional parameters consisting of FIR tap coefficients, clipping coefficients, and Vpp configurations output by GA or BP are alternately used in the IMDD system offline model. The PAM signal x input to the IMDD system offline model is passed through the system to obtain the output PAM signal Finally, the mean square error (MSE) of the input and output signals is calculated. The MSE is returned to the GA or BP, which then updates the multidimensional parameters based on the returned MSE. This process is repeated repeatedly to complete parameter optimization. Finally, the optimized parameters are deployed in the actual IMDD system to complete global optimization.

[0028] The specific implementation steps are as follows:

[0029] 1) Collect sample data: Build a Figure 2 In the real IMDD system shown, without adding an FIR module or clipping, a large number of random signals are generated using random numbers (limiting the signal values ​​to a feasible range) and input into the system. Different DAC peak-to-peak values ​​(Vpp) are then configured, and input and output test data between the DAC and the FFE are collected.

[0030] 2) Offline system modeling: Use the data tested in the previous step as training samples for the neural network and train the parameters of the neural network. After the training is completed, a neural network model corresponding to the system without FIR and Clipping can be obtained. A FIR filter and Clipping module are cascaded in front of the neural network model to obtain a complete offline IMDD system model. The multi-tap structure of the FIR filter and the structure of the neural network are as follows: Figure 2 As shown in Figure 1, these two modules have a multi-stage delay extraction structure before the signal stream input, where D represents the delay of one symbol period. The number of Ds can be determined based on the symbol length of the crosstalk. Multiple Ds indicate crosstalk between multiple symbols.

[0031] 3) Alternating optimization:

[0032] After establishing the offline IMDD system model, we first generate a population using the GA algorithm. Each individual in the population corresponds to a multidimensional parameter and is configured into the offline system model. Then, the PAM signal x is input into the offline IMDD system model. After the offline IMDD system model is completed, the PAM signal is output. The MSE performance of the offline IMDD system model under these multidimensional parameters is calculated and returned to the GA algorithm. Based on the returned MSE, the GA algorithm selects the three individuals with the best performance and performs selection, crossover, and mutation operations to generate a new population. The GA process is repeated five times before terminating. The multidimensional parameters corresponding to the optimal individuals are output as the GA-optimized multidimensional parameters and passed to the BP process as the initial parameter values.

[0033] In the BP process, the multi-dimensional parameters determined by GA are first configured into the offline IMDD system model, and then the PAM signal x is input into the offline IMDD system model. After the offline IMDD system model is completed, the PAM signal is output. The BP process calculates the MSE of the offline IMDD system model to determine whether it is stable. If not, backpropagation is performed to find the gradient and update the parameters. If not, the BP process is repeated until the MSE performance is stable. If so, the current multidimensional parameters are output as the multidimensional parameters for BP optimization, completing a GA-BP alternation.

[0034] After completing one GA-BP cycle, determine whether the current MSE performance has converged. If there is no significant improvement in performance, it is considered converged. If not, the multidimensional parameters optimized by BP are passed to the GA process. Otherwise, the process ends and the parameter values ​​are output as the optimal multidimensional parameters.

[0035] 4) Actual Testing: On the actual hardware IMDD system, the optimal multi-dimensional parameters of tap coefficients, clipping coefficients, and Vpp are assigned to the FIR filter, clipping, and DAC, respectively, completing the global optimization of the IMDD system. The IMDD system performance is then tested to determine whether it meets the system requirements.

Claims

1. A global optimization method for optical communication systems using a neural network combined with a genetic algorithm, characterized in that: Including steps: Steps to establish an offline optical communication system model: Determining a parameter module to be configured in an optical communication system; using multi-dimensional parameters to represent the parameter to be configured, where N parameters to be configured have N dimensions; and training a neural network to simulate a parameter module not to be configured in the optical communication system; The parameter module to be configured is placed before the trained neural network to form an offline optical communication system model; Multi-dimensional parameter alternating update steps: A genetic algorithm (GA) is used to update the multidimensional parameters in one round, wherein the received BP-optimized multidimensional parameters are used as the first individual in the initial population; the genetic algorithm (GA) outputs the multidimensional parameters corresponding to the optimal individual in the last population as the GA-optimized multidimensional parameters to the back propagation algorithm (BP); the GA-optimized multidimensional parameters are used as the BP initial parameter values, and BP is used to update the multidimensional parameters in one round, outputting the BP-optimized multidimensional parameters to complete a multidimensional parameter alternation update; Determine whether the currently updated multidimensional parameters meet the alternating update end condition. If so, output the latest multidimensional parameters as the multidimensional parameters for global optimization. Otherwise, repeat the alternating update of the multidimensional parameters. Actual configuration steps: Use the globally optimized multi-dimensional parameters to configure a real optical communication system to complete the global optimization of the optical communication system.

2. The method according to claim 1, wherein: The specific method of using genetic algorithm GA to update multidimensional parameters is as follows: generate an initial population, each individual in the population is a multidimensional parameter, and its value is randomly generated within the allowed range, and the first individual in the population is replaced by the multidimensional parameter optimized by BP; The multidimensional parameters corresponding to each individual in the population are sequentially configured into the offline optical communication system model, and the mean square error (MSE) of the input and output signals of the offline optical communication system model under each multidimensional parameter is calculated to reflect the fitness; the smaller the MSE, the higher the fitness; the three individuals with the highest fitness in the population are selected, crossover and mutation are performed with each other, and the new individuals obtained and the original three individuals form a new population; then the above operations of calculating fitness, selecting, crossover, mutating, and generating a new population are repeated five times; the multidimensional parameters corresponding to the optimal individual in the last population are output as the multidimensional parameters optimized by GA.

3. The method according to claim 1 or 2, wherein: When using the genetic algorithm GA to update the first round of multidimensional parameters, before receiving the multidimensional parameters optimized by BP, the multidimensional parameters corresponding to the first individual in the initial population are randomly generated within the allowed range.

4. The method according to claim 1, wherein: The specific method of using the back propagation algorithm BP to update multidimensional parameters is as follows: the multidimensional parameters optimized by GA are configured into the offline optical communication system model; the mean square error (MSE) of the model input and output signals under the current multidimensional parameter configuration is calculated, and then the gradient of the current multidimensional parameters is obtained through the back propagation method of the neural network according to the MSE, and then the multidimensional parameters are updated using the gradient descent method, and the updated multidimensional parameters are configured into the offline optical communication system model, and then the mean square error (MSE) of the model input and output signals under the current multidimensional parameter configuration is calculated, and the multidimensional parameter update operation is repeated until the BP iteration termination condition is reached, and the current multidimensional parameters are output as the multidimensional parameters optimized by BP.

5. The method according to claim 1, wherein: The optical communication system is an intensity modulation direct detection (IMDD) system; the parameters that need to be configured include FIR tap coefficients, clipping coefficients, and DAC peak-to-peak values ​​(Vpp); the parameter modules to be configured include an FIR filter, a clipping model, and a digital-to-analog converter (DAC); and the DAC peak-to-peak value (Vpp) is input into the neural network as a configurable parameter.

6. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

Images