Ultrafast pulse fiber transmission modeling method and system based on convolution feature extraction

Abstract

The present invention provides a method and system for modeling ultrafast pulse fiber transmission based on convolution feature extraction, which combines model-driven linear dispersion modeling with data-driven machine learning nonlinear feature modeling, can model strong nonlinear dynamic phenomena, and achieve pulse waveform prediction at different distances through iterative calculation. Furthermore, for nonlinear feature modeling of machine learning, a neural network structure combining convolution and local full connection is proposed, which realizes parameter sharing at different positions of the pulse, reduces the storage requirements of the model, and allows dynamic selection of high-energy parts of the pulse for nonlinear calculation. The entire modeling method has a certain degree of flexibility, and each module can be adjusted for different scenarios to achieve accurate modeling in different situations.

Description

Technical Field

[0001] The present invention relates to the technical field of optical fiber modeling, and in particular to an ultrafast pulse optical fiber transmission modeling method based on convolution feature extraction, and in particular to an ultrafast pulse optical fiber transmission modeling method based on convolution local feature extraction. Background Art

[0002] Ultrafast pulses propagate in optical fibers and are affected by both linear and nonlinear effects. The interaction between the two causes complex changes in the shape of the pulses, resulting in significant nonlinear dynamic evolution in the time and frequency domains. By modeling the pulse evolution process, the numerical results of the pulse at different times and distances can be quickly predicted, which can be applied to the extensive study of ultrafast optical pulses. The transmission of light waves in optical fibers can be accurately described by the nonlinear Schrödinger equation, and the step-by-step Fourier method is a widely used and relatively reliable numerical solution method.

[0003] A high-performance distributed optical fiber sensing system based on Epstein ultrafast pulse scanning is disclosed in the patent document with the publication number CN110160572A. When performing disturbance signal testing, the internal frequency conversion pulse light emitted by the Epstein ultrafast pulse scanning laser enters the sensing optical fiber through a circulator, and the backward Rayleigh scattered signal emitted by the sensing optical fiber enters the unbalanced Michelson interferometer through a coupler. By designing the arm length difference of the two interference arms, the backward Rayleigh scattered light at the position of the two adjacent interference arm length difference is interfered one by one. In this way, the received signal after passing through the unbalanced Michelson interferometer contains the phase difference signal caused by the external disturbance signal on the sensing optical fiber. Finally, the phase demodulation unit is used to demodulate the change of the relative phase difference signal over time, thereby realizing the dynamic measurement of the disturbance signal in the sensing optical fiber. At the same time, the use of the Epstein ultrafast pulse scanning laser can effectively realize its simultaneous monitoring of long distance, high frequency band signal and high spatial resolution.

[0004] With respect to the above-mentioned related technologies, the inventors believe that the step-by-step Fourier method is extremely computationally intensive and is difficult to be significantly accelerated through computational methods such as parallelization. Therefore, a new technical solution needs to be proposed. Summary of the invention

[0005] In view of the defects in the prior art, the object of the present invention is to provide a method and system for modeling ultrafast pulse fiber transmission based on convolution feature extraction.

[0006] According to the present invention, a method for modeling ultrafast pulse fiber transmission based on convolution feature extraction is provided, the method comprising the following steps:

[0007] Step S1: Collect input pulse data and pulse shapes with equally spaced steps Δz during the transmission process;

[0008] Step S2: shaping the continuous data collected in step S1 into an input data arrangement form that meets subsequent requirements;

[0009] Step S3: Modeling the optical fiber linear characteristics of the step length Δz according to the dispersion;

[0010] Step S4: The modeler uses a data-driven method based on a machine learning model to model unknown features, and uses convolutional units and fully connected units to achieve local nonlinear feature modeling;

[0011] Step S5: Compare the waveform outputted from step S4 with the reference waveform generated from step S2, calculate the error and perform gradient back propagation; the operations from step S2 to step S5 will be executed cyclically to reduce the error during training;

[0012] Step S6: After the training is completed, the model is used to model the pulse fiber transmission; the fiber input waveform is converted into normalized data, and steps S3 to S4 are iteratively executed to model the infinite length pulse fiber transmission.

[0013] Preferably, the data in step S1 comes from experimental measurement or numerical simulation; the input pulse data is recorded as Az(0), and the subsequent pulse waveforms at equally spaced step Δz positions are recorded as Az(1), Az(2), Az(3), ..., Az(M), where M is the serial number of the farthest position and each waveform is represented by N data points; this step is repeated multiple times to obtain multiple groups of data, each group being a waveform evolution, and the interval step Δz between different groups is the same.

[0014] Preferably, step S2 normalizes the data and inputs K adjacent pulses at the same time, that is, generates a K×N rectangular data block, where N is the number of points per pulse; when training the i-th step, the known waveforms of Az(i-K+1), Az(i-K+2),…, Az(i) are selected as the input of step S3, and the expected output is Az(i+1), and the data comes from the accurate data obtained in step S1.

[0015] Preferably, the linear modeler in step S3 completes the dispersion operation in the frequency domain; performs an N-point FFT transformation on the input waveform, multiplies different frequency points point by N known dispersion coefficients in the frequency domain, and performs an N-point IFFT transformation to obtain the dispersed waveform.

[0016] Preferably, step S4 uses a data-driven modeling method to model complex nonlinear phenomena that cannot be simply described; uses a multi-layer one-dimensional convolution and a multi-layer local full connection method to determine parameters through a machine learning gradient descent method; inputs K N-point pulses and outputs a single N-point pulse, and the model structure supports dynamic processing of different N values;

[0017] There are multiple cascaded one-dimensional convolutional layers and multiple cascaded local fully connected layers. P is the number of input points of the convolutional layer, which is dynamically adjusted and different from the number of pulse points N. The convolutional layer as a whole inputs K P-point pulses and outputs L P / M-point pulses. The overall input dimension of the local fully connected layer is L, and the output dimension is M. It is repeated on the P / M-point pulses and combined to obtain the P-point output.

[0018] Preferably, step S5 calculates the loss function value between the predicted value A'z(i) and the true value Az(i), and uses a back-propagation algorithm to calculate the gradient of the loss function for each undetermined parameter in step S4, and passes the gradient to the optimizer for parameter update.

[0019] Preferably, the step S6 takes the waveform of Az(0) repeated K times as input to obtain K×N data, and obtains the predicted value A'z(1) ​​after step S3 and step S4; then removes a pulse from the beginning of the input and splices the predicted output A'z(1) ​​obtained in the previous step at the end to obtain new K×N data, which is input into step S3 and step S4 again to obtain the predicted value A'z(2); and so on, iteratively calculate the subsequent step A'z(i);

[0020] Only some points with high energy are selected for nonlinear prediction, and the maximum amplitude of the input pulse at each moment is determined. If the maximum value of the pulse at a certain moment is greater than or equal to the threshold value T, it is marked; take P time points between the first mark and the last mark as the data to be nonlinearly predicted, and further input steps S3 and S4. After the remaining time points are input into step S3, step S4 is skipped and the value is copied to step S5.

[0021] The present invention also provides an ultrafast pulse optical fiber transmission modeling system based on convolution feature extraction, the system comprising the following modules:

[0022] Module M1: collects input pulse data and pulse shapes with equally spaced steps Δz during the transmission process;

[0023] Module M2: reshapes the continuous data collected by module M1 into an input data arrangement form that meets subsequent requirements;

[0024] Module M3: Modeling of optical fiber linear characteristics with step length Δz for dispersion;

[0025] Module M4: The modeler uses a data-driven system based on a machine learning model to model unknown features, using convolutional units and fully connected units to achieve local nonlinear feature modeling;

[0026] Module M5: Compare the waveform output by module M4 with the reference waveform generated by module M2, calculate the error and perform gradient backpropagation; the operations from module M2 to module M5 will be executed cyclically to reduce the error during training;

[0027] Module M6: After training, the model is used to model pulse fiber transmission; the fiber input waveform is converted into normalized data, and modules M3 to M4 are iteratively executed to model infinite-length pulse fiber transmission.

[0028] Preferably, the data in the module M1 is from experimental measurement or numerical simulation; the input pulse data is recorded as Az(0), and the subsequent pulse waveforms at equally spaced step Δz positions are recorded as Az(1), Az(2), Az(3), ..., Az(M), where M is the serial number of the farthest position, and each waveform is represented by N data points; the module is repeated multiple times to obtain multiple groups of data, each group is a waveform evolution, and the interval step Δz between different groups is the same;

[0029] The module M2 normalizes the data and inputs K adjacent pulses at the same time, that is, generates a K×N rectangular data block, where N is the number of points per pulse; when training the i-th step, the known waveforms Az(i-K+1), Az(i-K+2), …, Az(i) are selected as the input of the module M3, and the expected output is Az(i+1), and the data comes from the accurate data obtained by the module M1;

[0030] The linear modeler in the module M3 completes the dispersion operation in the frequency domain; performs N-point FFT transformation on the input waveform, multiplies different frequency points point by N known dispersion coefficients in the frequency domain, and performs N-point IFFT transformation to obtain the dispersed waveform.

[0031] Preferably, the module M4 uses a data-driven modeling system to model complex nonlinear phenomena that cannot be simply described; uses a multi-layer one-dimensional convolution and a multi-layer local full connection system to determine parameters through a machine learning gradient descent system; inputs K N-point pulses and outputs a single N-point pulse, and the model structure supports dynamic processing of different N values;

[0032] There are multiple cascaded one-dimensional convolutional layers and multiple cascaded local fully connected layers. P is the number of input points of the convolutional layer, which is dynamically adjusted and different from the number of pulse points N. The convolutional layer inputs K P-point pulses as a whole and outputs L P / M-point pulses. The overall input dimension of the local fully connected layer is L, and the output dimension is M. It is repeated on the P / M-point pulses and combined to obtain the P-point output.

[0033] The module M5 calculates the loss function value between the predicted value A'z(i) and the true value Az(i), and uses the back-propagation algorithm to calculate the gradient of the loss function for each undetermined parameter in the module M4, and passes the gradient to the optimizer for parameter update;

[0034] The module M6 takes the waveform of Az(0) repeated K times as input to obtain K×N data, and obtains the predicted value A'z(1) ​​after passing through modules M3 and M4; then removes a pulse from the beginning of the input and splices the predicted output A'z(1) ​​obtained in the previous step at the end to obtain new K×N data, which is input into modules M3 and M4 again to obtain the predicted value A'z(2); and so on, iteratively calculates the subsequent module A'z(i);

[0035] Only some points with high energy are selected for nonlinear prediction, and the maximum amplitude of the input pulse at each moment is determined. If the maximum value of the pulse at a certain moment is greater than or equal to the threshold value T, it is marked; P time points between the first mark and the last mark are taken as the data that need nonlinear prediction, and further input into modules M3 and M4. The remaining time points are input into module M3 and then module M4 is skipped, and the values ​​are copied to module M5.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention combines the existing precise dispersion modeling technology with machine learning technology, and proposes a modeling method for local nonlinearity; when designing the local nonlinear modeling method, a structure allowing dynamic length input is proposed to achieve efficient operation during prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0039] Figure 1 is a system structure diagram of the present invention;

[0040] Figure 2 It is a structural diagram of the nonlinear modeling part of step S4 and module M4 in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0042] Based on existing research, this invention combines linear and nonlinear feature separation and iterative modeling schemes to further optimize the accuracy and complexity of the fiber model. By combining the existing modeling method for the known effects of the fiber part with the neural network, a model that conforms to the transmission of ultrafast pulses in the fiber is designed. Compared with traditional numerical methods and other fiber modeling methods, the design scheme achieves significant performance improvement.

[0043] This embodiment discloses a method for modeling ultrafast pulse fiber transmission based on convolution feature extraction, which corresponds to the optical high-order soliton compression phenomenon and supercontinuum generation phenomenon.

[0044] To achieve the above object, the technical solution of the present invention is as follows: a method for modeling ultrafast pulse fiber transmission based on convolution feature extraction, the method comprising the following steps:

[0045] Step S1: Data acquisition, collect input pulse data and pulse shapes with equally spaced steps Δz during the transmission process; use the step-by-step Fourier method for numerical simulation. For the high-order soliton compression phenomenon, the number of points N = 1024, 3000 groups of pulses with different powers and pulse widths are input into one optical fiber, of which 2900 groups are used for training and 100 groups are used for testing. The number of simulation steps is 10,000 steps, and the training uses the pulse changes in 100 steps of Δz = 0.13m. For the supercontinuum generation phenomenon, the number of points N = 2048, 1300 groups of pulses with different powers are input into one optical fiber, of which 1250 groups are used for training and 50 groups are used for testing. The number of simulation steps is 20,000 steps, and the training uses the pulse changes in 200 steps of Δz = 0.1cm.

[0046] In step S1:

[0047] The data comes from experimental measurements or numerical simulations. The input pulse data is recorded as Az(0), and the waveforms of subsequent equally spaced step Δz positions are recorded as Az(1), Az(2), Az(3), …, Az(M), where M is the serial number of the farthest position and each waveform can be represented by N data points. This step can be repeated multiple times to obtain multiple sets of data, each of which is a waveform evolution, and the interval step Δz between different groups is the same. During training, each position under different input conditions will be iterated separately, so that a single model has the ability to handle different positions and different input conditions.

[0048] Step S2: Data normalization, shaping the continuous data collected in step S1 into an input data arrangement form that meets subsequent requirements; and step-cutting the input data into 15×N data arrangements.

[0049] In step S2:

[0050] Normalize the data into a rectangular data block of 15×N, where 15 is the number of pulses in the nonlinear step and N is the number of points per pulse. During training, the known waveforms of Az(i-14), Az(i-13), …, Az(i) are selected as the input of step S3, and the expected output is Az(i+1). The data comes from the accurate data obtained in step S1.

[0051] Step S3: Linear modeling, for dispersion, the optical fiber linear characteristic modeling with a step length Δz is performed; using the linear operation module in the step-by-step Fourier method, when the optical fiber parameters and Δz are known, the linear operation is performed on the input N-point pulse.

[0052] In step S3:

[0053] The linear modeler performs accurate dispersion operations in the frequency domain. The steps are: perform an N-point FFT transform on the input waveform, multiply each frequency point by N known dispersion coefficients in the frequency domain, and perform an N-point IFFT transform to obtain the dispersed waveform. The dispersion coefficient can be calculated based on the characteristics of the optical fiber using empirical knowledge. The beneficial effect of this step is that by directly incorporating frequency domain dispersion that can be simply and accurately modeled, the correlation between pulse points is reduced, thereby reducing the difficulty of modeling in subsequent steps at a relatively low cost.

[0054] Step S4: nonlinear modeling. This modeler uses a data-driven method based on a machine learning model to model unknown features. It uses convolutional units and fully connected units to achieve local nonlinear feature modeling. The value of M is 1. The input is complex data. Each complex number is regarded as two real numbers with real and imaginary parts. Therefore, the length of the input and output is recorded as P*2. Figure 2 The nonlinear modeling scheme shown in Figure 1 uses 3 convolutional layers with 90 channels each and 5 hidden fully connected layers with 90 output points each.

[0055] In step S4:

[0056] Use data-driven modeling methods to model complex nonlinear phenomena that cannot be accurately described. Use multi-layer one-dimensional convolution and multi-layer local full connection methods to determine parameters through machine learning gradient descent. Input 15 N-point pulses and output a single N-point pulse. The model structure supports dynamic processing of different N values.

[0057] In the nonlinear modeling step:

[0058] There are multiple cascaded one-dimensional convolutional layers and multiple cascaded local fully connected layers. P is the number of complex input points of the convolutional layer, which can be adjusted dynamically and is different from the number of pulse points N. The convolutional layer as a whole inputs 15 P*2 point real pulses and outputs 90 P point pulses. The overall input dimension of the local fully connected layer is exactly 90, and the output dimension is 2. It is repeated on the P point pulses, and the combination obtains P*2 point real output, that is, P point complex output. The beneficial effect of this structural design is that it allows the P point to change in integer multiples of M. For example, when M is 1 here, inputs of any length P are allowed, so that the nonlinearity inside the pulse of unknown length can be universally modeled. This allows only points with significant nonlinearity and higher energy to be selected as input during prediction, reducing the number of operations and improving the prediction speed.

[0059] Step S5: Parameter update, compare the waveform output from step S4 with the reference waveform generated in step S2, calculate the error and perform gradient back propagation. The operations from step S2 to step S5 will be executed cyclically to minimize the error during training; 120 rounds of training are performed, using a cosine annealing learning rate scheduler with hot restart, with the initial learning rate set to 0.0001, and the learning rate decay occurs every 1000 steps.

[0060] In step S5:

[0061] Calculate the loss function value between the predicted value A′z(i) and the true value Az(i), and use the back-propagation algorithm to calculate the gradient of the loss function for each undetermined parameter in step S4, and pass the gradient to the optimizer for parameter update. The training loss function uses the average relative value error, and the training round result with the smallest loss function on the test set is used as the final result. Through repeated parameter updates, the loss function value is made as small as possible, and the predicted value is as close to the true value as possible. The selection of the loss function and the selection of the optimizer can affect the effect of this step, and ultimately affect the effect of nonlinear modeling.

[0062] Step S6: After the training is completed, the model is used to model the pulse fiber transmission; the fiber input waveform is converted into normalized data, and steps S3 to S4 are iteratively executed to model the infinite length pulse fiber transmission.

[0063] In step S6:

[0064] Repeat the waveform of Az(0) 15 times as input to obtain 15×N data, and obtain the predicted value A′z(1) after steps S3 and S4. Then remove a pulse from the beginning of the input and splice the predicted output A′z(1) obtained in the previous step at the end to obtain new 15×N data, and input steps S3 and S4 again to obtain the predicted value A′z(2). Similarly, the subsequent step A′z(i) can be iteratively calculated. Only some points with higher energy can be selected for nonlinear prediction. The steps are to determine the maximum amplitude of the input pulse at each moment. If the maximum value of the pulse at a certain moment is greater than or equal to the threshold value T, mark it. Take the P time points between the first mark and the last mark as the data to be nonlinearly predicted, and further input S3 and S4. After the remaining time points are input into S3, S4 is skipped and the value is copied to step S5. The beneficial effect of this step is that the dynamic variable characteristic of the input pulse length in the aforementioned nonlinear modeling is utilized to reduce the amount of calculation during prediction.

[0065] The present invention also provides an ultrafast pulse optical fiber transmission modeling system based on convolution feature extraction, the system comprising the following modules:

[0066] Module M1: collects input pulse data and pulse shapes with equally spaced steps Δz during the transmission process; the data comes from experimental measurements or numerical simulations; the input pulse data is recorded as Az(0), and the subsequent pulse waveforms with equally spaced steps Δz are recorded as Az(1), Az(2), Az(3), …, Az(M), where M is the serial number of the farthest position and each waveform is represented by N data points; this module is repeated multiple times to obtain multiple groups of data, each group is a waveform evolution, and the interval step Δz between different groups is the same.

[0067] Module M2: reshapes the continuous data collected by module M1 into an input data arrangement format that meets subsequent requirements; normalizes the data and inputs 15 adjacent pulses at the same time, that is, generates a 15×N rectangular data block, where N is the number of points per pulse; when training the i-th step, selects the known waveforms of Az(i-14), Az(i-13),…, Az(i) as the input of module M3, and the expected output is Az(i+1), and the data comes from the accurate data obtained by module M1.

[0068] Module M3: For dispersion, the linear characteristics of the optical fiber with a step length of Δz are modeled; the linear modeler completes the dispersion operation in the frequency domain; an N-point FFT transformation is performed on the input waveform, and different frequency points are multiplied point by point by N known dispersion coefficients in the frequency domain, and an N-point IFFT transformation is performed to obtain the waveform after dispersion.

[0069] Module M4: The modeler uses a data-driven system based on a machine learning model to model unknown features, and uses convolutional units and fully connected units to achieve local nonlinear feature modeling; M is 1. The input is complex data, and each complex number is regarded as two real numbers with real and imaginary parts. Therefore, the length of the input and output is recorded as P*2, using Figure 2 The nonlinear modeling scheme shown in the figure. Three convolutional layers are used, each with 90 channels, and five hidden fully connected layers, each with 90 output points. A data-driven modeling system is used to model complex nonlinear phenomena that cannot be simply described; a multi-layer one-dimensional convolution and multi-layer local fully connected system is used, and the parameters are determined by a machine learning gradient descent system; 15 N-point pulses are input and a single N-point pulse is output, and the model structure supports dynamic processing of different N values; there are multiple cascaded one-dimensional convolutional layers and multiple cascaded local fully connected layers, P is the input complex number of the convolutional layer, which is dynamically adjusted and different from the number of pulse points N. The convolutional layer as a whole inputs 15 P*2-point real pulses and outputs 90 P-point pulses; the overall input dimension of the local fully connected layer is 90, and the output dimension is 2, which is repeated on the P-point pulses, and the combination obtains P*2-point real output, that is, P-point complex output.

[0070] Module M5: Compare the waveform output by module M4 with the reference waveform generated by module M2, calculate the error and perform gradient back propagation; the operations from module M2 to module M5 will be executed in a loop to reduce the error during training; calculate the loss function value between the predicted value A'z(i) and the true value Az(i), and use the back propagation algorithm to calculate the gradient of the loss function for each unknown parameter in module M4, and pass the gradient to the optimizer for parameter update.

[0071] Module M6: After training, the model is used to model pulse fiber transmission. The fiber input waveform is converted into normalized data, and modules M3 to M4 are iteratively executed to model infinite pulse fiber transmission. The waveform of Az(0) is repeated 15 times as input to obtain 15×N data, which is passed through modules M3 and M4 to obtain the predicted value A'z(1). Then, a pulse is removed from the beginning of the input and the predicted output A'z(1) ​​obtained in the previous step is concatenated at the end to obtain new 15×N data, which is input into modules M3 and M4 again to obtain the predicted value A'z(2). Similarly, the subsequent modules A'z(i) are iteratively calculated. Only some points with high energy are selected for nonlinear prediction, and the maximum amplitude of the input pulse at each moment is determined. If the maximum value of the pulse at a certain moment is greater than or equal to the threshold value T, it is marked. P time points between the first mark and the last mark are taken as data that need nonlinear prediction, which are further input into modules M3 and M4. The remaining time points are input into module M3 and then module M4 is skipped, and the values ​​are copied to module M5.

[0072] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0073] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for modeling ultrafast pulse fiber transmission based on convolution feature extraction, characterized in that: The method comprises the following steps: Step S1: Collect input pulse data and pulse shapes with equally spaced steps Δz during the transmission process; Step S2: shaping the continuous data collected in step S1 into an input data arrangement form that meets subsequent requirements; Step S3: Modeling the optical fiber linear characteristics of the step length Δz according to the dispersion; Step S4: The modeler uses a data-driven method based on a machine learning model to model unknown features, and uses convolutional units and fully connected units to achieve local nonlinear feature modeling; Step S5: Compare the waveform outputted from step S4 with the reference waveform generated from step S2, calculate the error and perform gradient back propagation; the operations from step S2 to step S5 will be executed cyclically to reduce the error during training; Step S6: After the training is completed, the model is used to model the pulse fiber transmission; the fiber input waveform is converted into normalized data, and steps S3 to S4 are iteratively executed to model the infinite length pulse fiber transmission.

2. The ultrafast pulse fiber transmission modeling method based on convolution feature extraction according to claim 1 is characterized in that: The data in step S1 comes from experimental measurement or numerical simulation; the input pulse data is recorded as Az(0), and the subsequent pulse waveforms with equally spaced step Δz positions are recorded as Az(1), Az(2), Az(3), ..., Az(M), where M is the serial number of the farthest position and each waveform is represented by N data points; this step is repeated multiple times to obtain multiple groups of data, each group is a waveform evolution, and the interval step Δz between different groups is the same.

3. The ultrafast pulse fiber transmission modeling method based on convolution feature extraction according to claim 1 is characterized in that: The step S2 normalizes the data and inputs K adjacent pulses at the same time, that is, generates a K×N rectangular data block, where N is the number of points per pulse; when training the i-th step, the known waveforms of Az(i-K+1), Az(i-K+2),…, Az(i) are selected as the input of step S3, and the expected output is Az(i+1), and the data comes from the accurate data obtained in step S1.

4. The ultrafast pulse fiber transmission modeling method based on convolution feature extraction according to claim 1 is characterized in that: The linear modeler in step S3 completes the dispersion operation in the frequency domain; performs an N-point FFT transformation on the input waveform, multiplies different frequency points point by N known dispersion coefficients in the frequency domain, and performs an N-point IFFT transformation to obtain the dispersed waveform.

5. The ultrafast pulse fiber transmission modeling method based on convolution feature extraction according to claim 1 is characterized in that: The step S4 uses a data-driven modeling method to model complex nonlinear phenomena that cannot be simply described; uses a multi-layer one-dimensional convolution and a multi-layer local full connection method to determine parameters through a machine learning gradient descent method; inputs K N-point pulses and outputs a single N-point pulse, and the model structure supports dynamic processing of different N values; There are multiple cascaded one-dimensional convolutional layers and multiple cascaded local fully connected layers. P is the number of input points of the convolutional layer, which is dynamically adjusted and different from the number of pulse points N. The convolutional layer as a whole inputs K P-point pulses and outputs L P / M-point pulses. The overall input dimension of the local fully connected layer is L, and the output dimension is M. It is repeated on the P / M-point pulses and combined to obtain the P-point output.

6. The ultrafast pulse fiber transmission modeling method based on convolution feature extraction according to claim 1 is characterized in that: The step S5 calculates the loss function value between the predicted value A'z(i) and the true value Az(i), and uses the back propagation algorithm to calculate the gradient of the loss function for each undetermined parameter in step S4, and passes the gradient to the optimizer for parameter update.

7. The ultrafast pulse fiber transmission modeling method based on convolution feature extraction according to claim 1 is characterized in that: The step S6 takes the waveform of Az(0) repeated K times as input to obtain K×N data, and obtains the predicted value A'z(1) ​​after step S3 and step S4; then removes a pulse from the beginning of the input and splices the predicted output A'z(1) ​​obtained in the previous step at the end to obtain new K×N data, which is input into step S3 and step S4 again to obtain the predicted value A'z(2); and so on, iteratively calculate the subsequent step A'z(i); Only some points with high energy are selected for nonlinear prediction, and the maximum amplitude of the input pulse at each moment is determined. If the maximum value of the pulse at a certain moment is greater than or equal to the threshold value T, it is marked; take P time points between the first mark and the last mark as the data to be nonlinearly predicted, and further input steps S3 and S4. After the remaining time points are input into step S3, step S4 is skipped and the value is copied to step S5.

8. An ultrafast pulse fiber transmission modeling system based on convolution feature extraction, characterized in that: The system includes the following modules: Module M1: collects input pulse data and pulse shapes with equally spaced steps Δz during the transmission process; Module M2: reshapes the continuous data collected by module M1 into an input data arrangement form that meets subsequent requirements; Module M3: Modeling of optical fiber linear characteristics with step length Δz for dispersion; Module M4: The modeler uses a data-driven system based on a machine learning model to model unknown features, using convolutional units and fully connected units to achieve local nonlinear feature modeling; Module M5: Compare the waveform output by module M4 with the reference waveform generated by module M2, calculate the error and perform gradient backpropagation; the operations from module M2 to module M5 will be executed cyclically to reduce the error during training; Module M6: After training, the model is used to model pulse fiber transmission; the fiber input waveform is converted into normalized data, and modules M3 to M4 are iteratively executed to model infinite-length pulse fiber transmission.

9. The ultrafast pulse fiber transmission modeling system based on convolution feature extraction according to claim 8, characterized in that: The data in the module M1 comes from experimental measurement or numerical simulation; the input pulse data is recorded as Az(0), and the subsequent pulse waveforms at equally spaced step Δz positions are recorded as Az(1), Az(2), Az(3), ..., Az(M), where M is the serial number of the farthest position and each waveform is represented by N data points; the module is repeated multiple times to obtain multiple groups of data, each group is a waveform evolution, and the interval step Δz between different groups is the same; The module M2 normalizes the data and inputs K adjacent pulses at the same time, that is, generates a K×N rectangular data block, where N is the number of points per pulse; when training the i-th step, the known waveforms Az(i-K+1), Az(i-K+2), …, Az(i) are selected as the input of the module M3, and the expected output is Az(i+1), and the data comes from the accurate data obtained by the module M1; The linear modeler in the module M3 performs dispersion operations in the frequency domain; Perform N-point FFT transformation on the input waveform, multiply different frequency points by N known dispersion coefficients point by point in the frequency domain, and perform N-point IFFT transformation to obtain the dispersed waveform.

10. The ultrafast pulse fiber transmission modeling system based on convolution feature extraction according to claim 8, characterized in that: The module M4 uses a data-driven modeling system to model complex nonlinear phenomena that cannot be simply described; uses a multi-layer one-dimensional convolution and multi-layer local full connection system, and determines parameters through a machine learning gradient descent system; inputs K N-point pulses and outputs a single N-point pulse, and the model structure supports dynamic processing of different N values; There are multiple cascaded one-dimensional convolutional layers and multiple cascaded local fully connected layers. P is the number of input points of the convolutional layer, which is dynamically adjusted and different from the number of pulse points N. The convolutional layer inputs K P-point pulses as a whole and outputs L P / M-point pulses. The overall input dimension of the local fully connected layer is L, and the output dimension is M. It is repeated on the P / M-point pulses and combined to obtain the P-point output. The module M5 calculates the loss function value between the predicted value A'z(i) and the true value Az(i), and uses the back-propagation algorithm to calculate the gradient of the loss function for each undetermined parameter in the module M4, and passes the gradient to the optimizer for parameter update; The module M6 takes the waveform of Az(0) repeated K times as input to obtain K×N data, and obtains the predicted value A'z(1) ​​after passing through modules M3 and M4; then removes a pulse from the beginning of the input and splices the predicted output A'z(1) ​​obtained in the previous step at the end to obtain new K×N data, which is input into modules M3 and M4 again to obtain the predicted value A'z(2); and so on, iteratively calculates the subsequent module A'z(i); Only some points with high energy are selected for nonlinear prediction, and the maximum amplitude of the input pulse at each moment is determined. If the maximum value of the pulse at a certain moment is greater than or equal to the threshold value T, it is marked; P time points between the first mark and the last mark are taken as the data that need nonlinear prediction, and further input into modules M3 and M4. The remaining time points are input into module M3 and then module M4 is skipped, and the values ​​are copied to module M5.

Images