Deep learning-based stimulated Brillouin scattering suppression method and system

Abstract

The invention relates to the technical field of fiber laser, in particular to a method and a system for suppressing stimulated Brillouin scattering by a spectral modulator driving waveform generation method based on deep learning, which can effectively counteract power amplification and nonlinear problems in a phase modulator, eliminate the influence of environmental factors on a modulation result, and improve the stability of the modulation result. According to the method, high-accuracy ideal spectrum modulation is realized, and meanwhile, the fine spectrum modulation effect can be better and better along with the increase of the data volume. And meanwhile, the control module of the high-power narrow-linewidth fiber laser system is subjected to integrated design and control, so that the miniaturization of the control module is realized, and the intelligence of the module is improved. Therefore, the problem that an existing stimulated Brillouin scattering suppression method cannot adapt to different optical systems and environment conditions is solved.

Description

Technical Field

[0001] The present invention relates to the field of fiber laser technology, and in particular to a method and system for suppressing stimulated Brillouin scattering based on deep learning. Background Art

[0002] The increasing application of high-power, narrow-linewidth fiber lasers in fields such as machining, earth science, laser weapons, nonlinear frequency conversion, and specialized applications is placing higher demands on laser quality and power. Due to issues such as nonlinear effects, optical damage, and thermal accumulation, increasing the power of single-frequency fiber lasers is becoming increasingly difficult. A single-frequency laser phase-modulated seed source, a time-domain stable, narrow-linewidth light source, is considered the optimal seed source technology for high-power, narrow-linewidth fiber lasers with a MOPA structure.

[0003] A narrow-linewidth seed light source is usually composed of a single-frequency light source and a phase modulation device. In order to suppress stimulated Brillouin scattering, a common method is to modulate the laser phase to achieve spectral broadening. Phase modulation methods include sinusoidal modulation, broadband noise modulation, pseudo-random code, designed waveforms, etc. Although current research shows that designed waveform modulation can obtain a denser and more uniform spectral broadening than other modulations and is more conducive to the suppression of stimulated Brillouin scattering, the generation of driving waveforms for traditional spectral modulators relies on empirical formulas or numerical simulations, which have the following problems: 1. Simulation and experimental deviation: The simulation model is difficult to fully simulate the nonlinear response and noise interference of actual optical devices; 2. Lack of physical constraints: Pure data-driven methods are prone to generate waveforms that do not conform to optical laws (such as Maxwell's equations and dispersion relations), resulting in hardware failure; 3. Low optimization efficiency: Hardware iteration requires repeated experiments, which is costly and has a long cycle.

[0004] In recent years, the rapid development of artificial intelligence (AI) has led to widespread research and application in various fields, particularly deep learning applications in optics, digital holography, and optical communications. Deep learning has provided new approaches to the field of optics. Deep learning-based computations have surpassed physics-based computations in signal processing and lightwave simulation in terms of speed and accuracy. Applying deep learning to waveform design and modulation spectrum can significantly improve modulation accuracy and avoid local optima. Deep neural networks can also significantly reduce workload, enabling rapid iteration and improvement. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for suppressing stimulated Brillouin scattering based on deep learning, aiming to solve the problem that existing stimulated Brillouin scattering suppression methods cannot adapt to different optical systems and environmental conditions.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for suppressing stimulated Brillouin scattering based on deep learning, comprising the following steps:

[0007] In a high-power narrow-linewidth fiber laser system, the laser source spectrum is modulated in a control module to obtain a modulation signal;

[0008] The modulation signal parameters are input into the constructed modulation model to obtain the ideal spectrum;

[0009] Inputting the modulated signal into an actual optical path to obtain a modulated spectrum;

[0010] The modulated signal and the modulated spectrum are input into a deep learning model for model training.

[0011] Input the spectrum into the constructed amplification system model to obtain the amplification power, and input the actual amplification power into the model to correct the amplification system;

[0012] The amplified power is input into the amplification system model to obtain the ideal spectrum;

[0013] Using ideal spectral parameters and a phase modulation algorithm based on deep learning, the modulation signal of the ideal spectrum is obtained;

[0014] The modulated signal is loaded into the phase modulator in the optical system to obtain an ideal spectrum shape, and is then sent to a subsequent amplification system.

[0015] The modulation signal parameters include modulation wavelength, modulation waveform, modulation bandwidth, modulation depth, modulation power and modulation phase.

[0016] In the second aspect, a stimulated Brillouin scattering suppression system based on deep learning adopts the stimulated Brillouin scattering suppression method based on deep learning described in the first aspect, including a control module and an optical path module, and the control module is connected to the optical path module.

[0017] Among them, the control module includes two temperature control, current drive, memory, waveform generation and control, adjustable filter, radio frequency amplifier, equalizer and digitally controlled attenuator; the two temperature control, current drive, memory and waveform generation and control modules are connected, and the digitally controlled attenuator, the adjustable filter, the radio frequency amplifier, the equalizer and the waveform generation and control are connected in sequence.

[0018] Among them, the optical path module includes a semiconductor pump source, a resonant cavity, a modulator, a coupler, a fine spectrometer 1, a circulator, an amplifier, a beam splitter 1, a beam splitter 2, a fine spectrometer 2, a power meter 1, a power meter 2 and a beam quality analyzer; the semiconductor pump source, the resonant cavity, the modulator, the coupler, and the fine spectrometer 1 are connected in sequence, the coupler, the circulator, the amplifier, the beam splitter 1, the beam splitter 2, and the fine spectrometer 2 are connected in sequence, the power meter 1 is connected to the circulator, the power meter 2 is connected to the beam splitter 1, and the beam quality analyzer is connected to the beam splitter 2.

[0019] The present invention's deep learning-based stimulated Brillouin scattering suppression method includes the following steps: modulating the laser source spectrum in a control module of a high-power narrow-linewidth fiber laser system to obtain a modulation signal; inputting the modulation signal parameters into a constructed modulation model to obtain an ideal spectrum; inputting the modulation signal into an actual optical path to obtain a modulated spectrum; and inputting the modulation signal and modulated spectrum into a deep learning model for model training. Inputting the spectrum into a constructed amplification system model to obtain amplified power, while simultaneously inputting the actual amplification power into the model to correct the amplification system; inputting the amplification power into the amplification system model to obtain an ideal spectrum; utilizing the ideal spectrum parameters, a modulation signal for the ideal spectrum is obtained through a deep learning-based phase modulation algorithm; and loading the modulation signal into a phase modulator in the optical path system to obtain an ideal spectrum shape, which is then fed into a subsequent amplification system. The present invention can effectively counteract nonlinear issues in the power amplification and phase modulator, eliminate the influence of environmental factors on the modulation results, and achieve highly accurate ideal spectrum modulation. Furthermore, this method can achieve increasingly better fine-tuned spectrum modulation effects as the amount of data increases. At the same time, the control module of the high-power narrow-linewidth fiber laser system was integrated with design and control, which not only achieved miniaturization of the control module but also improved its intelligence. This solved the problem that the existing stimulated Brillouin scattering suppression method could not adapt to different optical systems and environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 This is a flow chart of the stimulated Brillouin scattering suppression method based on deep learning provided by the present invention.

[0022] Figure 2This is a schematic diagram of the deep learning-based stimulated Brillouin scattering suppression system provided by the present invention.

[0023] In the figure: 1-temperature control, 2-current drive, 3-memory, 4-waveform generation and control module, 5-post-stage signal conditioner, 6-tunable filter, 7-RF amplifier, 8-equalizer, 9-digital controlled attenuator, 10-semiconductor pump source, 11-resonant cavity, 12-modulator, 13-coupler, 14-fine spectrometer 1, 15-circulator, 16-amplifier, 17-beam splitter 1, 18-beam splitter 2, 19-fine spectrometer 2, 20-power meter 1, 21-power meter 2, 22-beam quality analyzer. DETAILED DESCRIPTION

[0024] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0025] See also Figure 1 In a first aspect, the present invention provides a method for suppressing stimulated Brillouin scattering based on deep learning, comprising the following steps:

[0026] S1 modulates the laser source spectrum in the control module of the high-power narrow-linewidth fiber laser system to obtain a modulation signal;

[0027] Specifically, in the control module of the high-power narrow-linewidth fiber laser system, the spectrum of the laser source is first finely modulated. In this step S, the modulation signal parameters involved include modulation wavelength, modulation waveform, modulation bandwidth, modulation depth, modulation power, and modulation phase.

[0028] The optical equations for spectral modulation cover linear and nonlinear effects, time domain and frequency domain responses. Their mathematical form is complex but highly structured. In deep learning models, these equations can be converted into differentiable constraints through the transfer matrix method, numerical solutions of nonlinear differential equations, or parameterized physical models, so that both physical laws and data-driven characteristics can be taken into account when generating waveforms. In practical applications, appropriate equations must be selected for embedding based on the modulation and system complexity. Deep learning architecture with embedded physical constraints

[0029] Physical forward model: The optical transmission equation of the spectral modulator is embedded in the neural network as a constraint layer to ensure that the generated waveform conforms to physical laws;

[0030] Phase modulation

[0031]

[0032] Where V(t) is the modulation voltage, L is the crystal length, d is the electrode spacing, and n e is the refractive index of the crystal.

[0033] For optical system fiber links, spectral modulation is modeled by the transmission matrix,

[0034] Single-mode fiber transmission

[0035] E out (ω)=E in (ω)·e iβ(ω)L ·e -α(ω)L / 2

[0036] Where β(ω) is the propagation constant, α(ω) is the loss coefficient, and L is the fiber length.

[0037] Multilayer dielectric film filter:

[0038] The transmission matrix of each film layer is:

[0039]

[0040] in η j =n j cosθ j (TE polarization) or η j =n j / cosθ j (TM polarization). The total transmission matrix is ​​the product of the matrices of each layer:

[0041] M total =M N ·M N-1 …M N-1

[0042] In the Transformer-based generative network, the above equations can be embedded with physical constraints in the following way:

[0043] Transmission matrix constraint: If the generated waveform Y must satisfy M total Y=Y target , define the loss:

[0044]

[0045] Nonlinear effect constraints: By numerically solving the nonlinear Schrödinger equation (such as the split-step Fourier method), the difference between the generated waveform and the actual propagated waveform is calculated.

[0046] Hard constraints (architecture design):

[0047] Transfer Matrix Layer: Add a differentiable transfer matrix calculation layer after the decoder to force the output to satisfy

[0048]

[0049] Real-time SPM compensation: A nonlinear phase compensation module is inserted into the network to dynamically adjust the phase of the generated waveform.

[0050] S2 inputs the modulation signal parameters into the constructed modulation model to obtain the ideal spectrum;

[0051] The modulation signal parameters include modulation wavelength, modulation waveform, modulation bandwidth, modulation depth, modulation power and modulation phase.

[0052] Specifically, based on the modulation parameters set in step S1, these parameters are input into the constructed modulation model. Through the operation of the model, an ideal spectral shape can be predicted and generated.

[0053] S3 inputs the modulated signal into the actual optical path to obtain a modulated spectrum;

[0054] Specifically, the modulation signal set in step S1 is input into the actual optical path system. Through the modulator 12 in the optical path, the spectral form after modulation can be observed, which is a key step in converting theoretical predictions into practical operations.

[0055] S4 inputs the modulated signal and the modulated spectrum into a deep learning model for model training.

[0056] Rapid model convergence is achieved through a two-stage strategy of pre-training with simulation data and fine-tuning with experimental data.

[0057] Specifically, to further improve the accuracy and efficiency of spectral modulation, the modulation signal and observed spectral data are input into a deep learning model. Through model training and learning, the modulation parameters can be continuously optimized and adjusted to achieve an effect closer to the ideal spectrum.

[0058] Training process:

[0059] Step 1: Generate 100,000 sets of “drive waveform-spectral response” pairing data through transfer matrix simulation;

[0060] In addition, the actual 100,000 sets of "input drive waveform-spectral response" are used as correction data;

[0061] Step 2: The data processor performs cropping preprocessing on the collected spectral data through the acquisition module, adjusts the center position based on the effective position of the collected spectrum, sets parameters according to the spectrum, intercepts all the information of the spectrum, removes the background noise, and reduces the interference of noise;

[0062] Step 3: The acquisition module performs data enhancement on the pre-processed data, so that the limited data produces an effect equivalent to more data and then stores it in groups;

[0063] Step 4: The processing module builds a deep learning CNN network model based on the stored waveform data characteristics. The convolutional neural network consists of an input layer, a hidden layer, and an output layer. The nodes of each layer are interconnected with all the nodes of the next layer. The data enters the convolutional neural network from the input layer and propagates forward to the hidden layer in turn until it reaches the output layer.

[0064] Step 5: Input the waveform data after data enhancement into the CNN network model for training to complete model optimization and achieve matching selection between waveform data and spectrum, including the following steps:

[0065] Step 51: Initialize the model parameter weights of the convolutional neural network;

[0066] Step 52: Input the enhanced waveform data into the neural network, and propagate forward through the convolutional layer, pooling layer, and fully connected layer in the hidden layer to obtain the output value;

[0067] Step 53: Calculate the error between the output value of the convolutional neural network and the target value;

[0068] Step 54: Compare the calculated error with the expected error value set by initialization. When the error is greater than the expected error value set by initialization, the error is transmitted back to the neural network from the output layer, and the errors of the fully connected layer, the pooling layer, and the convolution layer are obtained in sequence. The sum of the errors of each layer is the total error of the neural network. The weights are updated according to the obtained errors, and the process returns to step 52 for forward propagation and continuous training. Otherwise, the training is terminated to obtain a model for CNN waveform generation after optimization, and the neural network connection weights are optimized. The optimal network model parameters obtained after multiple optimizations are output as a pth file and stored in the analysis module.

[0069] In step 2, the circle() function in Python language is used for clipping preprocessing

[0070] In step 3, the transforms() function set is applied to perform data enhancement on the preprocessed training set data.

[0071] The construction of the deep learning CNN network model in step 4 requires analyzing the waveform characteristics and building it from simple to complex according to different modulation modes. For simple waveforms with obvious feature distinctions, a shallower neural network is used to save training time; for complex waveforms with vague feature distinctions, a deeper neural network is used to improve accuracy.

[0072] In step 4, the CNN network is built using the Pytorch framework.

[0073] The spectral data of the sample with known predicted values ​​is input into the convolutional neural network model to train the model weights. After multiple rounds of training, an optimal model is obtained to obtain the trained model.

[0074] Step 3): Input the spectral data of the unknown predicted value of the sample into the trained model, and output the predicted value result of the waveform data.

[0075] The convolutional neural network model is mainly composed of an input layer, a convolution layer (1), a convolution layer (2), a convolution layer (3), a stretching layer, a fully connected layer and an output layer connected in sequence; the original full-band spectrum curve is input into the input layer;

[0076] In the step 1, specifically:

[0077] 1.1) The convolutional neural network model is mainly composed of an input layer, a convolutional layer (1), a convolutional layer (2), a convolutional layer (3), a stretching layer, a fully connected layer, and an output layer connected in sequence; the original full-band spectrum curve is input into the input layer;

[0078] The first convolutional layer contains a convolution module using 10 convolution kernels, all of which have the same size;

[0079] The second convolutional layer uses three parallel modules: two convolution modules and one pooling module. The output of the first convolutional layer is input into two convolution modules and one pooling module respectively. Each convolution module uses a convolution kernel. The convolution kernels of the two convolution modules are different. Each convolution module contains four 1×1×10 convolution kernels. The pooling module contains four parallel maximum pooling structures.

[0080] The third convolutional layer uses four convolution modules, and the four convolution modules use four different convolution kernels. The first convolution module contains four 1×1×10 first convolution kernels, the second convolution module contains four m×1×4 second convolution kernels, the third convolution module contains four n×1×4 third convolution kernels, and the fourth convolution module contains four 1×1×4 fourth convolution kernels. m and n represent the lengths of the second and third convolution kernels, respectively. The output of the first convolution kernel is input to the first convolution module of the third convolution layer, and the two convolution modules and one pooling module of the second convolution layer are respectively input to the last three convolution modules of the third convolution layer; the stretching layer stretches the output of the third convolution layer into a one-dimensional feature vector;

[0081] The ranges of the sizes and step sizes of different convolution kernels in the above three convolution layers are as follows: the convolution kernel size range in the first convolution layer is 2-19, and the convolution kernel step range is 2-9; the length of the first convolution kernel in the second convolution layer is set to 1, and the step size of the first convolution kernel is set to 1, the length of the second convolution kernel in the second convolution layer is set to 1, and the step size of the second convolution kernel is set to 1; the length of the first convolution kernel in the third convolution layer is set to 1, the length m of the second convolution kernel in the third convolution layer is in the range of 2-5, the length n of the third convolution kernel in the second convolution layer is in the range of 6-9, and the step size range of the four convolution kernels in the third convolution layer is 2-9.

[0082] S5 inputs the spectrum into the constructed amplification system model to obtain the amplification power, and at the same time inputs the actual amplification power into the model to correct the amplification system;

[0083] Specifically, after obtaining the modulated spectrum, its amplification performance is evaluated. The spectrum is input into the constructed amplification system model. Based on the morphology and characteristics of the spectrum, the model predicts the amplified power level. The amplification system model is then modified based on the actual amplified power, ultimately resulting in a model that conforms to the actual amplification system. This step S5 provides an important reference for subsequent spectral amplification.

[0084] S6 inputs the amplified power into the amplification system model to obtain the ideal spectrum;

[0085] Specifically, based on the amplified power predicted in step S5, the amplification system model is again used to generate an ideal spectral shape after amplification. This spectral shape not only takes into account the power amplification but also the morphology and characteristics of the spectrum, and serves as a reference target for subsequent spectrum optimization and modulation.

[0086] S7 uses ideal spectral parameters and a phase modulation algorithm based on deep learning to obtain a modulation signal with an ideal spectrum.

[0087] Specifically, to achieve a more precise and ideal spectral morphology, a deep learning-based phase modulation algorithm is utilized. By inputting the desired spectral parameters, the algorithm automatically calculates the modulation signal, specifically the phase modulation signal, required to achieve this desired spectrum. This step combines deep learning with traditional spectral modulation techniques, achieving intelligent optimization and precise control of spectral morphology.

[0088] S8 loads the modulated signal into the phase modulator in the optical system to obtain an ideal spectrum shape, and sends it to the post-amplification system.

[0089] Specifically, finally, the modulated signal calculated in step S7 is applied to the phase modulator in the optical path system. Through precise adjustment of the phase modulator, a spectral shape highly consistent with the ideal spectral form can be obtained. This precisely modulated and optimized spectral shape is then fed into the subsequent amplification system to achieve stable output of the high-power, narrow-linewidth fiber laser. This step S is the final embodiment of spectral modulation and optimization and a key step in achieving improved performance in high-power, narrow-linewidth fiber laser systems.

[0090] See also Figure 2 In the second aspect, a stimulated Brillouin scattering suppression system based on deep learning adopts the stimulated Brillouin scattering suppression method based on deep learning described in the first aspect, including a control module and an optical path module, and the control module is connected to the optical path module.

[0091] The control module includes two temperature control modules 1, a current drive module 2, a memory module 3, a waveform generation and control module 4, a post-stage signal conditioner 5, an adjustable filter 6, a radio frequency amplifier 7, an equalizer 8 and a digitally controlled attenuator 9; the two temperature control modules 1, the current drive module 2, the memory module 3 and the post-stage signal conditioner 5 are respectively connected to the waveform generation and control module 4, and the digitally controlled attenuator 9, the adjustable filter 6, the radio frequency amplifier 7, the equalizer 8 and the post-stage signal conditioner 5 are connected in sequence.

[0092] The optical path module includes a semiconductor pump source 10, a resonant cavity 11, a modulator 12, a coupler 13, a fine spectrometer 14, a circulator 15, an amplifier 16, a spectrometer 17, a spectrometer 2 18, a fine spectrometer 2 19, a power meter 1 20, a power meter 21 and a beam quality analyzer 22; the semiconductor pump source 10, the resonant cavity 11, the modulator 12, the coupler 13 and the fine spectrometer 14 are connected in sequence, the coupler 13, the circulator 15, the amplifier 16, the spectrometer 17, the spectrometer 2 18 and the fine spectrometer 2 19 are connected in sequence, the power meter 1 20 is connected to the circulator 15, the power meter 2 21 is connected to the spectrometer 1 17, and the beam quality analyzer 22 is connected to the spectrometer 2 18.

[0093] In this implementation, different initial parameter combinations are applied to the modulator 12, and spectral data from the modulator 12 output is collected through extensive experiments. The input data (modulator 12 parameters) and output data (spectrum) are standardized. A reinforcement learning model is used to adjust the parameters in real time to adapt to varying environmental conditions. The spectral error is used as a reward function, and the mean square error (MSE) is used to calculate the difference between the current spectral output and the target spectrum. This error is fed back into the model, and the modulator 12 parameters are adjusted through the reinforcement learning algorithm's reward mechanism, gradually optimizing the modulator 12 parameters until the output spectrum matches the target spectrum.

[0094] The semiconductor pump source 10 and the resonant cavity 11 serve as the output of a single-frequency laser source, and are spectrally modulated by the modulator 12. The broadened optical signal is divided into two paths by a 10:90 coupling beam splitter. The weak light is connected to a spectrometer with an accuracy of 2pm. The collected spectral data is input into the AI ​​software system. The strong light passes through a circulator 15, an amplifier 16, a beam splitter 17, and a beam splitter 2 18. The C end of the circulator 15 is connected to a power meter for measuring the return light power. The beam splitter 17 splits the light into two beams. The strong light is introduced into the power meter to test the amplified power. The beam splitter 2 18 introduces the two beams into a fine spectrometer and a beam quality analyzer 22 respectively.

[0095] In this system, we first use experimental data for offline training to establish a basic model, collect real-time spectral data for feature extraction, and use multi-layer neural networks to learn the deep information of the data. Through deep learning, spectrum perception is achieved and useful bands are identified. Subsequently, online training is used to continuously adjust to changes in the actual environment (i.e., model adjustment).

[0096] Beneficial effects:

[0097] 1. The technical solution provided by the present invention adopts AI-based intelligent phase modulation, which can effectively offset the nonlinear problems in power amplification and phase modulators, eliminate the influence of environmental factors on the modulation results, and achieve high-accuracy ideal spectral modulation. At the same time, this method can achieve better and better fine spectral modulation effects as the amount of data increases.

[0098] 2. The present invention integrates the design and control of the control module of the high-power narrow-linewidth fiber laser system, which not only realizes the miniaturization of the control module but also improves the intelligence of the module.

[0099] 3. The present invention has strong generalization and supports the generation of driving waveforms for multiple types of modulators (such as electro-optic modulators and acousto-optic modulators); at the same time, it has low hardware costs, and experimental-simulation collaborative training reduces the number of hardware iterations by 50%.

[0100] The above disclosure is merely a preferred embodiment of the deep learning-based stimulated Brillouin scattering suppression method and system of the present invention. It is certainly not intended to limit the scope of the present invention. A person skilled in the art will understand that all or part of the processes of the above embodiments and equivalent changes made in accordance with the claims of the present invention still fall within the scope of the invention.

Claims

1. A deep learning-based stimulated Brillouin scattering suppression method, characterized by: The following steps are involved: In a high-power narrow-linewidth fiber laser system, the laser source spectrum is modulated in a control module to obtain a modulation signal; The modulation signal parameters are input into the constructed modulation model to obtain the ideal spectrum; Inputting the modulated signal into an actual optical path to obtain a modulated spectrum; The modulated signal and the modulated spectrum are input into a deep learning model for model training. Input the spectrum into the constructed amplification system model to obtain the amplification power, and input the actual amplification power into the model to correct the amplification system; The amplified power is input into the amplification system model to obtain the ideal spectrum; Using ideal spectral parameters and a phase modulation algorithm based on deep learning, the modulation signal of the ideal spectrum is obtained; The modulated signal is loaded into the phase modulator in the optical system to obtain an ideal spectrum shape, and is then sent to a subsequent amplification system.

2. The method for suppressing stimulated Brillouin scattering based on deep learning according to claim 1, wherein: The modulation signal parameters include modulation wavelength, modulation waveform, modulation bandwidth, modulation depth, modulation power and modulation phase.

3. A deep learning-based stimulated Brillouin scattering suppression system, comprising the deep learning-based stimulated Brillouin scattering suppression method according to any one of claims 1 to 2, characterized in that: It includes a control module and an optical path module, and the control module is connected to the optical path module.

4. The deep learning-based stimulated Brillouin scattering suppression system according to claim 3, wherein: The control module includes two temperature control modules, a current drive module, a memory module, a waveform generation and control module, a post-stage signal conditioner, an adjustable filter, a radio frequency amplifier, an equalizer, and a digitally controlled attenuator; the two temperature control modules, the current drive module, the memory module, and the post-stage signal conditioner are respectively connected to the waveform generation and control module, and the digitally controlled attenuator, the adjustable filter, the radio frequency amplifier, the equalizer, and the post-stage signal conditioner are connected in sequence.

5. The deep learning-based stimulated Brillouin scattering suppression system according to claim 4, wherein: The optical path module includes a semiconductor pump source, a resonant cavity, a modulator, a coupler, a fine spectrometer 1, a circulator, an amplifier, a first beam splitter, a second beam splitter, a second fine spectrometer 2, a first power meter, a second power meter and a beam quality analyzer; the semiconductor pump source, the resonant cavity, the modulator, the coupler and the first fine spectrometer are connected in sequence, the coupler, the circulator, the amplifier, the first beam splitter, the second beam splitter and the second fine spectrometer are connected in sequence, the first power meter is connected to the circulator, the second power meter is connected to the first beam splitter, and the beam quality analyzer is connected to the second beam splitter.

Images