Reverse design method of dissipative soliton resonance laser based on machine learning

Abstract

The invention discloses a dissipative soliton resonance laser reverse design method based on machine learning, and the method comprises the steps: 1) building two groups of associated data sets which are respectively used for training an SVM and a BP neural network, the SVM is used for predicting the dissipative soliton resonance pulse convergence, and the BP neural network is used for predicting the output time domain and spectral waveform of a laser; (2) a PSO algorithm particle swarm is initialized, the position parameter of each particle corresponds to a set of random laser cavity parameters, an SVM model is input in advance, particles meeting the convergence condition are screened out, and the particles serve as an initial particle swarm again; 3) iteratively updating particle position information, inputting the particle position information into the BP neural network, and predicting pulse time domain and spectral waveform information; and when the mean square error of the predicted waveform and the set target waveform is lower than a set value, outputting a group of optimal laser cavity parameters to realize the reverse design of the laser. The problem that a dissipative soliton resonance fiber laser is difficult to design under the condition of low-absolute-value net cavity dispersion is effectively solved.

Description

Technical Field

[0001] The present invention belongs to the field of laser technology and artificial intelligence, and specifically relates to a method for predicting pulse characteristics of a Figure-9 dissipative soliton resonant fiber laser and a laser reverse design method based on machine learning. Background Art

[0002] High-energy passively mode-locked fiber lasers have been widely used in fiber-optic communications, fiber-optic sensing, material processing, medical diagnosis, national defense, precision measurement, and scientific research. Among them, due to the clamping effect of peak power, the pulse energy and duration of dissipative soliton resonant pulses increase linearly with the increase of pump power, while the peak power remains constant without light wave splitting, which is an effective way to achieve high-energy single pulses. Such pulses can exist in normal and anomalous dispersion fiber laser cavities, but with the decrease of absolute net cavity dispersion, the stable mode-locking area is significantly reduced. In general, small net cavity dispersion means relatively high peak power, which limits the output of high-peak power dissipative soliton resonant pulses to a certain extent. In addition, the pulse time domain presents a rectangular structure and a rectangular structure with a central narrow peak in the low absolute value normal and anomalous dispersion regions, respectively, accompanied by a single peak and a double peak structure in the spectrum. Therefore, studying the characteristics of dissipative soliton resonant pulses in the low dispersion range and designing mode-locked fiber lasers in a targeted manner are of great significance for achieving high-energy, high-peak power pulse output and studying complex pulse dynamics.

[0003] At present, in the numerical simulation of mode-locked fiber lasers, it is usually necessary to use the split-step Fourier method (SSFM) to solve the generalized nonlinear Schrödinger equation (GNLSE) to obtain the mapping relationship between the output pulse characteristics and the input cavity parameters, which requires traversing the unknown laser cavity parameter space. Obviously, the higher the dimension of the traversed cavity parameter space, the more universal the conclusions obtained. However, when the cavity parameter dimension exceeds three dimensions, the traditional laser solution method of using SSFM to solve GNLSE will consume a lot of computing resources and simulation time, resulting in low computational efficiency. In addition, when the pulse characteristics are known, the traditional simulation method cannot obtain the corresponding cavity parameters through inverse calculation, which has great limitations.

[0004] In recent years, machine learning technology has been gradually applied to the numerical simulation of mode-locked lasers. By learning large-scale data, it can efficiently establish the mapping relationship between pulse characteristics and cavity parameters, thereby greatly reducing the calculation time and avoiding the tedious calculation process of traditional methods. Machine learning combined with optimization algorithms also plays a key role in laser reverse design. However, existing numerical studies on pulse characteristics and pulse dynamics using machine learning mainly focus on relatively low-energy traditional solitons, dispersion-managed solitons, and dissipative soliton lasers. There are still relatively few numerical studies on high-energy, high-peak-power dissipative soliton resonant lasers, especially for 8- or 9-shaped cavity lasers with low absolute net cavity dispersion and complex cavity structures. Summary of the invention

[0005] In view of the above problems, the present invention provides a method for reverse design of low absolute value cavity dispersion dissipative soliton resonant laser based on machine learning. The method combines machine learning and optimization algorithms, and can effectively predict the time domain and frequency domain waveforms of dissipative soliton resonant pulses in high-dimensional parameter space, study the pulse stability region, the change of pulse characteristic parameters with cavity parameters, and realize the reverse design of dissipative soliton resonant pulse lasers.

[0006] The core of the present invention is to use machine learning algorithms to pre-judge pulse convergence and train a model that can simultaneously predict the dissipative soliton resonance pulse waveform information in the normal and anomalous dispersion regions. Finally, combined with a selected optimization algorithm, the reverse design of the low absolute net cavity dispersion Figure-9 laser is achieved. The specific process is as follows: Figure 1 , the corresponding specific technical solutions are as follows:

[0007] 1. Construct a correlation data set of laser cavity parameters and their corresponding pulse characteristics.

[0008] for Figure 2 The dissipative soliton resonant fiber laser shown in Figure-9 (it can also be a dissipative soliton resonant laser of other structures) first uses the traditional method of SSFM combined with GNLSE to simulate the transmission of the light beam in the laser cavity, and then constructs a related data set containing a large number of laser cavity parameters and their corresponding pulse convergence and pulse characteristics (including time domain and spectral waveform information). Among them, the laser cavity parameters and their corresponding pulse convergence labels are used as samples to train the SVM model to pre-judge whether the laser is stable to the dissipative soliton resonant pulse state. Subsequently, under the convergence condition, the laser cavity parameters and the corresponding pulse waveform information are used as samples to train the BP neural network to predict the time domain and spectral waveform of the pulse.

[0009] The present invention only focuses on the characteristics of strongly stable dissipative soliton resonance pulses, and other pulse types, unstable dissipative soliton resonance pulse states, and pulse non-convergence states are all classified as non-convergence of dissipative soliton resonance pulses.

[0010] Taking into account the steep time domain edge characteristics of the rectangular dissipative soliton resonant pulse, and the characteristics that the pulse presents a narrow peak in the time domain under the condition of anomalous dispersion and has a low-intensity sidelobe between two main peaks, the present invention adopts an adaptive sampling method for sampling to ensure the accuracy of the predicted waveform. When ignoring high-order dispersion and high-order nonlinearity, the time domain and spectral waveforms of the pulse are usually symmetrical about the simulation window, that is, symmetrical about T = 0s and ω = 0THz, respectively, so only half of the time domain and frequency domain data can be collected to save calculation time, and at the same time ensure that the central peak area point of the pulse under the condition of anomalous dispersion is effectively collected.

[0011] 2. Train the support vector machine (SVM) model to pre-judge the convergence of dissipative soliton resonant pulses.

[0012] Under the condition of low absolute net cavity dispersion, the region where the dissipative soliton resonant pulse exists stably is significantly reduced. Therefore, it is particularly important to pre-judge whether the laser converges to a stable dissipative soliton resonant pulse state to save computing resources and time.

[0013] Preferably, SVM, as a commonly used binary classification model in machine learning, is particularly suitable for solving classification problems in high-dimensional parameter space. In the present invention, by taking the laser cavity parameters in the sample for training the SVM model established in step 1 as input, and taking the convergence of the dissipative soliton resonance pulse (convergence or non-convergence) as the label output of the classification problem, the SVM model can find the optimal hyperplane according to the training data and separate different convergence categories, showing strong generalization ability.

[0014] The kernel function determines the quality of SVM model training. Therefore, in order to deal with nonlinear problems under multi-parameter conditions, the present invention comprehensively compares the characteristics of linear kernel, polynomial kernel, radial basis (RBF) kernel and sigmoid kernel, and uses the RBF kernel function that can be mapped to infinite dimensions to capture the nonlinear relationship between cavity parameters and convergence. In addition, during the model training process, cross-validation and grid search techniques are used to optimize the hyperparameters of SVM.

[0015] 3. Train a back-propagation (BP) neural network that can map laser cavity parameters to pulse temporal and spectral waveforms.

[0016] After completing the SVM model training in step 2, the present invention further verifies the convergence of the samples used to train the BP neural network in step 1 using the SVM model to ensure that all samples can converge to the dissipative soliton resonance pulse state. Subsequently, the sample data set is divided into a training set, a validation set, and a test set according to 70%, 15%, and 15%. The training set is used for training the BP neural network, the validation set is used for model optimization, and the test set is used for final verification, thereby establishing a BP neural network model that can efficiently predict the time domain and spectral waveforms of the dissipative soliton resonance pulse.

[0017] The input layer nodes of the BP neural network correspond to the laser cavity parameters. The number of hidden layers and nodes are optimized according to the cross-validation method. The output layer nodes contain the sampling point intensity, maximum sampling time and spectrum range of the pulse time domain and spectral waveform. For example, when the sampling points in the time domain and frequency domain are 128 and 128 respectively, the output layer nodes are 258. The mean square error (MSE) is used as the loss function to optimize the cavity parameters during the training process.

[0018] 4. The BP neural network combined with the particle swarm optimization (PSO) algorithm completes the reverse design of the dissipative soliton resonance fiber laser.

[0019] The reverse design of the present invention is to search for the global optimal solution of the laser cavity parameters by adopting the PSO algorithm based on the given target time domain and spectral waveforms to optimize the BP neural network output, and then reversely derive the optimal set of laser cavity parameters.

[0020] First, the particle swarm is randomly initialized, and the position parameters of each particle (corresponding to a set of random cavity parameters) are input into the SVM model trained by step 2 to determine the pulse convergence, and the particles that meet the pulse convergence conditions are screened out as the initial particle swarm. Subsequently, the position parameters of the converged particles are sequentially input into the BP neural network trained by step 3 to complete multiple sets of pulse time domain and spectral waveform predictions. Further, the PSO algorithm is used to globally optimize the particle swarm, gradually adjust the laser cavity parameters and input them into the BP neural network for pulse waveform prediction, until the pulse waveform predicted by the BP neural network is closest to the given target pulse waveform, that is, the MSE value is the smallest. At this point, the PSO algorithm iteration stops and outputs a set of optimal laser cavity parameters, thereby realizing the reverse design of the dissipative soliton fiber laser.

[0021] The advantages of the present invention are as follows:

[0022] 1. Based on the machine learning algorithm, the present invention can quickly and accurately predict the convergence, time domain and spectral waveform characteristics of dissipative soliton resonant pulses in high-dimensional cavity parameter space, as well as the variation law of pulse parameters with cavity parameters. Compared with the traditional laser solution method based on SSFM to solve GNLSE, this method significantly improves the computational efficiency and saves computing resources.

[0023] 2. This invention is the first to study dissipative soliton resonant lasers under low absolute net cavity dispersion conditions based on machine learning. When the net cavity dispersion transitions from normal to anomalous dispersion, the time domain waveform of the dissipative soliton resonant pulse transitions from a rectangular shape to a rectangular shape with a central peak, and the spectrum transitions from a single peak to a double peak structure. Based on the trained BP neural network, the dynamic change of pulse characteristics with cavity parameters can be quickly studied.

[0024] 3. Using the PSO algorithm to optimize the trained BP neural network can quickly realize the reverse design of lasers with multi-dimensional cavity parameters, effectively solving the problem of narrow convergence region and relatively difficult design of dissipative soliton resonant lasers under low absolute net cavity dispersion conditions, greatly saving development time and R&D costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Design a flow chart for the overall system.

[0026] Figure 2 Figure-9 is the schematic diagram of the dissipative soliton resonance fiber laser.

[0027] Figure 3 Schematic diagram of pulse time domain and spectral waveform sampling.

[0028] Figure 4 Figure-9 is the flow chart of the reverse algorithm for dissipative soliton resonance fiber laser. DETAILED DESCRIPTION

[0029] The following is combined with Figure 1 The present invention is further described in detail. The examples given are only used to explain the present invention but not to limit the scope of the present invention.

[0030] 1. The present invention first uses the traditional laser simulation method combining SSFM and GNLSE to Figure 2 The dissipative soliton resonant laser shown in Figure-9 is simulated to obtain the convergence, time domain and spectral waveform information of the dissipative soliton resonant pulse under the condition of multi-dimensional laser cavity parameter changes. The convergence label will be used for the training of the SVM model, and the corresponding time domain and spectral waveform information under the convergence condition will be used for the subsequent BP neural network training.

[0031] Figure 2Figure-9 is the schematic diagram of the dissipative soliton resonant fiber laser. The laser cavity includes a pump source 1, a wavelength division multiplexer (WDM) 2, an erbium-doped fiber (EDF) 3, a non-reciprocal phase shifter (PS) 4, a dispersion compensating fiber (DCF) 5, a single-mode fiber (SMF) 6, a coupler (OC) 7, a tunable filter (TF) 8, a single-mode fiber (SMF) 9, a dispersive delay line (DDL) 10 and a reflector 11.

[0032] There are 8 adjustable cavity parameters involved in the present invention, namely, the length of EDF (L EDF ) and gain saturation energy (E sat )、DCF length (L DCF ), SMF 9 length (L SMF ), PS phase shift The coupling ratio (k) of OC, the bandwidth (Δλ) of TF and the dispersion (Δβ2) of DDL are defined in the upper and lower limits of each parameter. The value of an input variable in a single sample is a random number within its definition domain, which is randomly generated by a computer program. Finally, 5000 samples were prepared for training the SVM and BP neural network models. In order to study the characteristics of dissipative soliton resonant pulses under low absolute net cavity dispersion conditions, the appropriate fiber length and Δβ2 range were selected before simulation to ensure that the net cavity dispersion was kept at -1.5 to 1.5ps. 2 research within the scope.

[0033] In the simulation, the traditional laser simulation method of SSFM combined with GNLSE is used to simulate the transmission of the light field in the laser cavity, and the maximum number of cycles is set to 3000. If within 3000 cycles, the relative energy change of the pulses in two adjacent cycles has already been reduced to 10 -9 If the pulse intensity curves in both time and frequency domains show smooth characteristics, the dissipative soliton resonance pulse is considered to have converged and is marked as 1. On the contrary, if the pulse does not meet the above conditions, it is considered to have not converged and is marked as -1. The convergence mark is used for subsequent SVM model training.

[0034] Based on the steep pulse edge characteristics of the rectangular dissipative soliton resonance pulse and the unique time-frequency domain characteristics of the pulse in the anomalous dispersion region, the present invention adopts an adaptive sampling method for sampling. The pulse waveform is normalized before sampling to eliminate the influence of the peak intensity on the neural network. Subsequently, the waveform gradient is calculated on the time domain waveform to determine the regions of the rising edge, peak, and falling edge of the pulse, and sampling points are added in these regions; for the spectral waveform, the gradient is calculated using the same method, focusing on sampling regions where the frequency components change dramatically. In order to speed up the calculation and better reflect the waveform information at the peak, half of the waveform is sampled, such as Figure 3 As shown, the red area is the key sampling area. The sampling time and frequency range correspond to the maximum intensity It,max and I f,max Reduce to I t,max / 200 and I f,max / 200, the corresponding horizontal coordinate value is recorded as ΔT max and Δf max To accurately restore the pulse waveform, ΔT max and Δf max The combined time domain and spectral sampling intensity values ​​are used as sample points of the BP neural network.

[0035] 2. The present invention divides the laser cavity parameters obtained in step 1 and their corresponding pulse convergence data sets into a training set and a test set in a ratio of 80:20, where 80% of the data is used for model training and 20% of the data is used for model verification. Each sample contains a set of laser cavity parameters (L EDF 、E sat , L DCF , L SMF , k, Δλ and Δβ2) and the convergence label (1 or -1) of the dissipative soliton resonant pulse.

[0036] Subsequently, the SVM model is trained using the training set, and the SVM model is optimized by maximizing the distance between the support vector and the hyperplane, thereby achieving accurate classification of convergence labels. During the training process, the RBF kernel function is used to process the nonlinear characteristics of the sample data, and the RBF kernel function is tuned using grid search technology combined with cross-validation method. The specific tuning parameters are the kernel parameter σ and the penalty parameter C. σ controls the width of the kernel function, and C balances the fitting of the training set and the generalization ability of the model.

[0037] After the SVM model is trained, its performance is evaluated using the test set data. The evaluation indicators include accuracy, precision, recall, and F1-score. The closer these indicators are to 1, the better the performance of the SVM model, thus ensuring the reliability and generalization ability of the model in judging the convergence of dissipative soliton resonance pulses.

[0038] 3. After completing the SVM model training in step 2, the present invention uses the model to further verify the sample convergence used to train the BP neural network in step 1 to ensure that all 5000 samples can converge to the dissipative soliton resonant pulse state. Subsequently, the samples are divided into a training set, a validation set, and a test set at a ratio of 70%, 15%, and 15%. Each sample includes a set of laser cavity parameters (L EDF 、E sat , L DCF , L SMF , k, Δλ and Δβ2) and their corresponding pulse time domain and spectral waveform sampling point intensity, sampling time range ΔTmax and frequency range Δf max .

[0039] Subsequently, the BP neural network was trained using the training set. By learning the input laser cavity parameters, the mapping relationship between the laser cavity parameters and the pulse time domain and spectral waveforms was learned. The input layer nodes of the BP neural network correspond to 8 laser cavity parameters; the hidden layer is used to capture the nonlinear relationship between the cavity parameters and the pulse waveform. Each layer uses the ReLU activation function to alleviate the problem of gradient disappearance during training; the output layer is the pulse time domain and spectral waveform information obtained by the adaptive sampling method. The number of nodes L = m + n + 2, where m and n are the number of sampling points of the time domain and frequency domain waveforms, respectively.

[0040] During the training process, the mean square error (MSE) is used as the Loss function. Because the prediction accuracy of the time domain and spectral waveforms must be considered at the same time, the Loss function is the sum of the two Loss functions, that is:

[0041]

[0042] Among them, MSE time is the MSE of the time domain waveform, MSE freq is the MSE of the spectral waveform, y t,i is the intensity value of the i-th time domain sampling point, is the corresponding target waveform intensity value, y f,i is the intensity value of the i-th frequency domain sampling point, is the corresponding target spectral intensity value. In addition, the optimization function selects the Adam optimizer to adaptively adjust the learning rate, and the initial learning rate is set to 0.001.

[0043] After completing the model training, the regression graphs of the training set, validation set, and test set are drawn to visually evaluate the consistency between the model output results and the target value, and the correlation coefficient R is used. 2 To quantify the model performance. 2 The closer it is to 1, the better the model training effect is. In addition, within the defined range of cavity parameter variations, multiple groups of cavity parameter combinations can be randomly selected for prediction and further verified using traditional numerical methods.

[0044] 4. The present invention combines the trained BP neural network with the PSO algorithm to achieve the reverse design of the dissipative soliton resonant fiber laser. Compared with the commonly used genetic algorithm (GA), the PSO algorithm is faster and more accurate because it does not have the "crossover" and "mutation" operations of GA. The reverse algorithm flow chart is as follows Figure 4 shown.

[0045] (1) Initialize the particle swarm in the PSO algorithm, including the number of particles, particle position, velocity, learning factor, inertia factor, and maximum number of iterations.

[0046] In the initialization stage of the PSO algorithm, the present invention introduces the SVM model to screen the randomly generated particle swarm to ensure that the position parameters (L EDF 、E sat , L DCF , L SMF , k, Δλ and Δβ2) meet the convergence conditions of dissipative soliton resonant pulses. This operation can greatly reduce the computational overhead caused by invalid particles, narrow the search space and increase the computational speed.

[0047] The particle velocity v controls the moving direction and step length of the particle in the parameter space. The initial value is set to 0 or randomly generated within a certain range. The learning factors c1 and c2 are set to the empirical value 2. The inertia factor w is a non-negative number and adopts a linear decreasing strategy, as shown in formula (2), where w max and w min are the maximum and minimum inertia weights, which are set to 0.9 and 0.4 respectively, to ensure the global exploration capability in the early stage of the search and enhance the local convergence capability in the later stage, thereby improving the accuracy and convergence speed of the global optimal solution. max is the maximum number of iterations, and t is the current number of iterations.

[0048]

[0049] (2) Determine the BP neural network structure and input the position of each particle into the trained BP neural network in turn, thereby obtaining the sampling points of the pulse time domain and frequency domain and restoring the time domain and spectral waveforms.

[0050] (3) Calculate the fitness function f(x) of each particle j ), that is, calculate the MSE value:

[0051]

[0052] Where j represents the jth particle and its value range is 1 to N; n and m represent the number of sampling points in the time domain and spectrum, respectively. This operation improves the physical accuracy and robustness of the search results.

[0053] (4) After each iteration, determine whether the termination condition is reached. Termination condition A: The global optimal fitness function maxf(x j ) is less than the preset target accuracy value. Termination condition B: The number of iterations is greater than the set maximum number of iterations. If the above termination conditions A or B are met, the global optimal particle position G is output best,d , otherwise update the individual optimal position Pbest,d and the global optimal position G best,d , update the position and velocity of each particle according to formula (4). is the speed after the tth iteration in the d-dimensional parameter space, is the position after the tth iteration in the d-dimensional parameter space, and r1 and r2 are independent random numbers between (0,1).

[0054]

[0055] Subsequently, the updated particle swarm data is input into the BP neural network again, and steps 2-4 are repeated until the set termination condition is met. Through dynamic adjustment and intelligent search, the present invention effectively balances the performance of global search and local convergence, ensures that the cavity parameter combination that meets the design requirements is quickly found, and provides strong guidance and reference value for experimental design.

[0056] Although the specific embodiments of the present invention are disclosed for the purpose of illustration, the purpose is to help understand the content of the present invention and implement it accordingly, those skilled in the art will understand that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the best embodiment, and the scope of the present invention is subject to the scope defined in the claims.

Claims

1. A method for reverse design of a dissipative soliton resonant laser based on machine learning, comprising the following steps: 1) establishing a correlation data set, the correlation data set comprising: (1) a first correlation data set of laser cavity parameters and their corresponding dissipative soliton resonance pulse convergence labels; (2) a second correlation data set of laser cavity parameters and pulse time domain and spectral waveform information under convergence conditions; 2) using the samples of the first associated data set to train the SVM model to obtain a classification model that can predict the convergence of the dissipative soliton resonance pulse, which is used to pre-judge whether the laser converges to a stable dissipative soliton resonance pulse state; 3) using the samples in the second associated data set to train a BP neural network to obtain a BP neural network capable of mapping laser cavity parameters with pulse time domain and spectral waveforms; 4) Initialize the particle swarm in the PSO algorithm, set the position parameters of each particle to a set of random laser cavity parameters, and input them into the classification model trained in step 2) to predict the pulse convergence corresponding to the output, and select the particles that meet the convergence conditions as the initial particle swarm of the optimization algorithm; 5) Use the PSO optimization algorithm to globally optimize the particle swarm, iteratively adjust and update the laser cavity parameters, and input each set of updated laser cavity parameters into the BP neural network trained in step 3) in turn to output the predicted pulse waveform information; when the pulse waveform predicted and output by the BP neural network meets the preset conditions and the set target pulse waveform, Stop iteration and output a set of optimal laser cavity parameters; 6) Design a dissipative soliton fiber laser according to a set of optimal laser cavity parameters obtained in step 5).

2. The method according to claim 1, characterized in that SSFM is used to solve the transmission of the GNLSE simulated light beam in the laser and establish the associated data set. Combined with the dissipative soliton resonant pulse characteristics under low absolute net cavity dispersion conditions, the waveform is sampled using an adaptive sampling method to ensure data density in key areas. The time domain and frequency domain waveforms are symmetrical about the time and frequency axes, respectively, and computing resources are saved by collecting half of the time domain and frequency domain data.

3. The method according to claim 1 or 2, characterized in that: When the first associated data set is used to train the SVM model, the laser cavity parameters in the sample are input into the SVM model, and the SVM is supervised and trained in combination with the label information to obtain a classification model that can predict the convergence of dissipative soliton resonance pulses.

4. The method according to claim 1 or 2, characterized in that: When the second associated data set is used to train the BP neural network, the laser cavity parameters in the sample are used as the BP neural network input, and the time domain and spectral waveform data of the sample are used as the output to establish a BP model for predicting pulse waveforms.

5. The method according to claim 4, characterized in that The input layer nodes of the BP neural network correspond to the laser cavity parameters, and the number of hidden layers and nodes are tuned according to the cross-validation method; the output layer nodes include the sampling point intensity, maximum sampling time and spectrum range of the pulse time domain and spectral waveform; the mean square error is used as the loss function to optimize the laser cavity parameters during the training process, and Adam is used as the optimizer.

6. The method according to claim 1, characterized in that The preset condition is that the mean square error between the pulse waveform predicted and output by the BP neural network and the target pulse waveform is the smallest.

7. The method according to claim 1, characterized in that The dissipative soliton resonance laser is a Figure-9 dissipative soliton resonance fiber laser, a semiconductor laser or a solid laser.

8. A server, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images