High-power fiber laser seed source spectrum broadening optimization method and system

Abstract

The present invention provides a high-power fiber laser seed source spectrum broadening optimization method and system, including: step S1: parameter initialization, designing a driving electrical signal according to a target spectrum; step S2: outputting a driving electrical signal, broadening the seed source spectrum; step S3: measuring and storing the broadened seed source spectrum information, parsing the spectrum information to obtain an empirical value and storing it; step S4: training and updating the optimization network according to the empirical value to obtain a maximum benefit value; step S5: establishing a corresponding model between the driving electrical signal and the seed source spectrum, and optimizing the driving electrical signal spectrum corresponding to the target spectrum. Based on conventional phase modulation mathematical relations and experimental data, the present invention uses a deep learning algorithm to establish a physical model between the driving electrical signal and the seed source broadening spectrum. Through a customized driving signal design, precise control of each frequency signal can be achieved, effectively improving the control accuracy and flexibility of the seed source spectrum.

Description

Technical Field

[0001] The present invention relates to the field of high-power continuous fiber lasers, and in particular to a method and system for optimizing the spectral broadening of a seed source of a high-power fiber laser. Background Art

[0002] High-power continuous fiber lasers have the advantages of high beam quality, good reliability, high electro-optical conversion efficiency, small size and convenient thermal management. They have important applications in coherent synthesis, nonlinear frequency conversion, industrial processing and military attack and defense. All-fiber lasers based on the master oscillator power amplifier (MOPA) structure are easy to implement and have significant power improvement performance. They are considered to be the preferred structure for high-power output of fiber lasers. However, when the pump power is too high, the laser output will be affected by a variety of nonlinear effects in the fiber, including cross-phase modulation, stimulated Raman scattering effect, stimulated Brillouin scattering (SBS) effect and four-wave mixing. Among them, the stimulated Brillouin scattering (SBS) effect in the fiber has the lowest threshold, which becomes a bottleneck problem that limits the maximum output power of the laser. Common methods for suppressing the SBS effect include: increasing the effective mode field area of ​​the fiber, changing the Brillouin gain spectrum, reducing the effective length of the fiber, and phase modulation to broaden the seed source line width. Taking into account the complexity of the manufacturing process, the degree of improvement in the SBS effect threshold and operability, the scheme of using phase modulation to broaden the seed source bandwidth has the characteristics of simple and controllable operation and high power enhancement efficiency, and has become a research hotspot for suppressing the SBS effect in high-power fiber lasers.

[0003] In high-power continuous fiber lasers, the modulation signals for phase modulation spectrum broadening of seed sources mainly include sinusoidal signals, white noise sources (White Noise Source, WNS), pseudo-random binary sequences (Pseudo-Random Binary Sequence, PRBS) and multi-frequency signals. The current driving electrical signal and the spectrum after broadening have poor controllability, the broadened spectrum is basically Gaussian, and the self-pulsation phenomenon caused by the randomness of the signal limits the maximum output power. Under the same bandwidth and total power, seed sources with different spectral shapes have different laser output power thresholds. However, it is still a mystery what kind of seed source spectrum has the best beam quality and the highest nonlinear effect suppression efficiency. In addition, the model exploration and ideal spectrum optimization research on the optimal driving signal of the seed source are relatively lacking. Therefore, the present invention aims to establish a model between the seed source driving electrical signal and the broadened spectrum shape, design and optimize the phase modulator driving electrical signal under the condition of known seed source target spectrum type, so as to obtain the ideal target spectrum type, guide the spectral design direction of the seed source for different application scenarios, and improve the laser power efficiency.

[0004] Patent document CN113852424A (application number: CN202111117160.0) discloses a seed source spectrum broadening method, device and high-power continuous fiber laser system. The seed source spectrum broadening method includes: outputting a polarization-maintaining narrow linewidth continuous laser signal through a seed source; generating a pseudo-random bit sequence signal as a driving signal through a driving module and outputting it to a power amplifier module; the power amplifier module performs power amplification processing on the driving signal output by the driving module; the phase modulator receives the polarization-maintaining narrow linewidth continuous laser signal output by the seed source and the driving signal output by the power amplifier module, and broadens the polarization-maintaining narrow linewidth continuous laser signal by high-order phase modulation. The present invention establishes a physical model between the driving electrical signal and the seed source broadening spectrum. When the target spectrum of the seed source is known, the phase modulator driving electrical signal is designed and optimized to obtain the ideal target spectrum, and the precise control of each frequency signal is achieved through the custom driving signal design. However, the invention simply broadens the seed source spectrum, and does not achieve the broadening of different target spectra of the seed source for different application scenarios.

[0005] Patent document CN113991409A (application number: CN202111264892.2) discloses a seed source spectrum broadening system and method based on binary multi-frequency signal drive, including: the seed source outputs a narrow linewidth optical signal; designs a binary multi-frequency signal and outputs it through an arbitrary waveform generator; after filtering out the signal harmonics, power amplification is performed, and then the frequency doubling components are filtered out; the seed source spectrum is broadened by phase modulation; the shape of the seed source spectrum after broadening is measured and recorded; after the seed source passes through an optical isolator, it passes through an optical power pre-amplifier; the seed source signal light is connected to the first optical circulator port, and is led out from the second optical circulator port to the optical power main amplifier, and an optical power meter is connected to the third optical circulator port; the output signal of the second optical circulator port is detected in real time by an optical power meter. Although the invention optimizes the spectral details within a limited spectral broadening range, it does not achieve the broadening of the target spectrum of the seed source. At the same time, the technical problem that factors including the nonlinear response of the device will cause spectral spectrum errors, the invention does not solve this existing technical problem. Summary of the invention

[0006] In view of the defects in the prior art, the object of the present invention is to provide a method and system for optimizing the spectral broadening of a high-power fiber laser seed source.

[0007] A high-power fiber laser seed source spectrum broadening optimization method provided by the present invention comprises:

[0008] Step S1: parameter initialization, designing the driving electrical signal according to the target spectrum;

[0009] Step S2: outputting a driving electrical signal to broaden the spectrum of the seed source;

[0010] Step S3: measuring and storing the broadened seed source spectrum information, analyzing the spectrum information to obtain empirical values ​​and storing them;

[0011] Step S4: training and updating the spectrum optimization network according to the empirical value to obtain the maximum benefit value;

[0012] Step S5: Establish a corresponding model between the driving electric signal and the seed source spectrum, and optimize the driving electric signal spectrum corresponding to the target spectrum.

[0013] Preferably, in step S1:

[0014] Initialize the driving electrical signal, agent, experience playback pool, spectral state and define the state transition process. According to the target spectral shape of the seed source, design the driving electrical signal of the corresponding spectral type.

[0015] Initialization objects include Actor network μ θ , Critic Network Q w, Target Actor Network TargetCritic Network and experience replay pool;

[0016] Initialize the Actor network μ with random parameters θ, w θ and Critic Network Q w ;

[0017] Using random parameters Initialize the Target Actor Network and Target Critic Network

[0018] Among them, setting The initial value of is equal to θ, set The initial value is equal to w; the format of experience storage in the experience replay pool is:

[0019] E=(s i ,a i ,r i ,s i+1 ,done)

[0020] Among them, s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed;

[0021] The spectral design of the driving electrical signal includes amplitude, phase, frequency interval, bandwidth and amplitude envelope shape; the spectral information of the seed source includes amplitude, bandwidth, in-band flatness and amplitude envelope shape.

[0022] Preferably, in step S2:

[0023] The signal generator outputs a preset driving electrical signal; the generated driving electrical signal is input to the RF input port of the phase modulator after power amplification, and the seed source outputs a laser with a preset wavelength, which is input to the phase modulator for spectrum modulation and broadening;

[0024] The signal generator includes an arbitrary waveform generator and a digital-to-analog converter;

[0025] The phase modulation spectrum broadening process of the seed source is single-level phase modulation or multi-level phase modulation;

[0026] In step S3:

[0027] The spectrometer measures and stores the broadened seed source spectrum information;

[0028] Initialize the seed source spectrum state:

[0029] s 0 =(1,I limit (1),0,0,…0)

[0030] The i-th state of the seed source spectrum is defined as:

[0031] s i =(id,I limit (id),I 1 ,I 2 ,…I m )

[0032] Among them, id represents the index of the position to be determined, and the positions are sorted in the order of frequency components from 1 to m. limit (id) represents the maximum power limit of the idth position, I id ,id=1,2,…m represents the power value of the idth frequency component;

[0033] The seed source spectrum state transfer process is defined as:

[0034] From the initial state s 0 =(1,I limit (1),0,0,…0), the agent generates action a 0 =I 1 , set the power value of the frequency component and get reward r 0 =0, update the state to s 1 =(2,I limit (2),I 1 ,0,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0035] From the spectral state s i =(i+1,I limit (i+1),I 1 ,I 2 ,…I i ,…0) starts, Agent generates action a i =I i+1 , set the power value of the i+1th frequency component, and then get the reward r i =0, update the state to si+1 =(i+2,I limit (i+2),I 1 ,I 2 ,…I i ,I i+1 ,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0036] And so on, until the state is updated to the end state s m =(0,0,I 1 ,I 2 ,…I m ), set the state transfer end flag done = True, indicating that the entire state transfer process is over. When done = True, let r 0 =r 1 =…=r i =…r m =reward, reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed.

[0037] Preferably, in step S4:

[0038] Train and update the agent based on the experience gained; repeat the training and strategy exploration until the maximum benefit value is obtained;

[0039] The state transfer process E=(s i ,a i ,r i ,s i+1 , done) as experience, stored in the experience replay pool, and repeat the state update operation until the specified amount of experience is generated;

[0040] Among them, s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed;

[0041] Minimize the loss according to the following formula:

[0042]

[0043] Loss2=-Q w (s i ,μ θ (s i )),

[0044] Reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed;

[0045] represents the target Q network parameter, and Loss1 represents the Critic network Q w Optimization formula, Loss2 represents the Actor network μ θ Optimization formula, MSE means mean square error;

[0046] Update Critic Network Q w and Actor Network μ θ , Q w (s i ,a i ) represents the critic network Q w Facing state s i and action is a i Output at time, μ θ (s i ) represents the Actor network μ θ Facing state s i Output action, Q w (s i ,μ θ (s i )) represents the critic network Q w Facing state s i and action is μ θ (s i ), α is the discount factor.

[0047] Preferably, in step S5:

[0048] The corrected driving electric signal and the seed source broadened spectrum are analyzed, the corresponding model between the driving electric signal and the seed source spectrum is reconstructed, and the driving electric signal spectrum corresponding to the target spectrum is obtained through an iterative optimization method;

[0049] Repeat the above training update operation, and the training end judgment standard is: the absolute value of the difference between the current reward value and the previous reward value is less than χ 1 , the absolute value of the difference between the current Agent output action and the last Agent output action is less than χ 2 , where χ 1 ≥0, χ 2≥0 is the set decision threshold, and finally the driving signal spectrum is determined according to the convergence value of the Agent output action.

[0050] A high-power fiber laser seed source spectrum broadening optimization system provided by the present invention comprises:

[0051] Module M1: parameter initialization, designing driving electrical signals according to the target spectrum;

[0052] Module M2: outputs driving electrical signals to broaden the spectrum of the seed source;

[0053] Module M3: measure and store the broadened seed source spectrum information, analyze the spectrum information to obtain the empirical value and store it;

[0054] Module M4: Train and update the spectrum optimization network based on empirical values ​​to obtain the maximum benefit value;

[0055] Module M5: Establish a corresponding model between the driving electrical signal and the seed source spectrum, and optimize the driving electrical signal spectrum corresponding to the target spectrum.

[0056] Preferably, in the module M1:

[0057] Initialize the driving electrical signal, agent, experience playback pool, spectral state and define the state transition process. According to the target spectral shape of the seed source, design the driving electrical signal of the corresponding spectral type.

[0058] Initialization objects include Actor network μ θ , Critic Network Q w , Target Actor Network TargetCritic Network and experience replay pool;

[0059] Initialize the Actor network μ with random parameters θ, w θ and Critic Network Q w ;

[0060] Using random parameters Initialize the Target Actor Network and Target Critic Network

[0061] Among them, setting The initial value of is equal to θ, set The initial value is equal to w; the format of experience storage in the experience replay pool is:

[0062] E=(s i ,ai ,r i ,s i+1 ,done)

[0063] Among them, s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed;

[0064] The spectral design of the driving electrical signal includes amplitude, phase, frequency interval, bandwidth and amplitude envelope shape; the spectral information of the seed source includes amplitude, bandwidth, in-band flatness and amplitude envelope shape.

[0065] Preferably, in the module M2:

[0066] The signal generator outputs a preset driving electrical signal; the generated driving electrical signal is input to the RF input port of the phase modulator after power amplification, and the seed source outputs a laser with a preset wavelength, which is input to the phase modulator for spectrum modulation and broadening;

[0067] The signal generator includes an arbitrary waveform generator and a digital-to-analog converter;

[0068] The phase modulation spectrum broadening process of the seed source is single-level phase modulation or multi-level phase modulation;

[0069] In the module M3:

[0070] The spectrometer measures and stores the broadened seed source spectrum information;

[0071] Initialize the seed source spectrum state:

[0072] s 0 =(1,I limit (1),0,0,…0)

[0073] The i-th state of the seed source spectrum is defined as:

[0074] s i =(id,I limit (id),I 1 ,I 2 ,…I m )

[0075] Among them, id represents the index of the position to be determined, and the positions are sorted in the order of frequency components from 1 to m. limit (id) represents the maximum power limit of the idth position, I id ,id=1,2,…m represents the power value of the idth frequency component;

[0076] The seed source spectrum state transfer process is defined as:

[0077] From the initial state s 0 =(1,I limit (1),0,0,…0), the agent generates action a 0 =I 1 , set the power value of the frequency component and get reward r 0 =0, update the state to s 1 =(2,I limit (2),I 1 ,0,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0078] From the spectral state s i =(i+1,I limit (i+1),I 1 ,I 2 ,…I i ,…0) starts, Agent generates action a i =I i+1 , set the power value of the i+1th frequency component, and then get the reward r i =0, update the state to s i+1 =(i+2,I limit (i+2),I 1 ,I 2 ,…I i ,I i+1 ,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0079] And so on, until the state is updated to the end state s m =(0,0,I 1 ,I 2 ,…I m ), set the state transfer end flag done = True, indicating that the entire state transfer process is over. When done = True, let r 0 =r 1 =…=r i =…r m=reward, reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed.

[0080] Preferably, in the module M4:

[0081] Train and update the agent based on the experience gained; repeat the training and strategy exploration until the maximum benefit value is obtained;

[0082] The state transfer process E=(s i ,a i ,r i ,s i+1 , done) as experience, stored in the experience replay pool, and repeat the state update operation until the specified amount of experience is generated;

[0083] Among them, s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed;

[0084] Minimize the loss according to the following formula:

[0085]

[0086] Loss2=-Q w (s i ,μ θ (s i )),

[0087] Reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed;

[0088] represents the target Q network parameter, and Loss1 represents the Critic network Q w Optimization formula, Loss2 represents the Actor network μ θ Optimization formula, MSE means mean square error;

[0089] Update Critic Network Q w and Actor Network μ θ , Q w (s i,a i ) represents the critic network Q w Facing state s i and action is a i Output at time, μ θ (s i ) represents the Actor network μ θ Facing state s i Output action, Q w (s i ,μ θ (s i )) represents the critic network Q w Facing state s i and action is μ θ (s i ), α is the discount factor.

[0090] Preferably, in the module M5:

[0091] The corrected driving electric signal and the seed source broadened spectrum are analyzed, the corresponding model between the driving electric signal and the seed source spectrum is reconstructed, and the driving electric signal spectrum corresponding to the target spectrum is obtained through an iterative optimization method;

[0092] Repeat the above training update operation, and the training end judgment standard is: the absolute value of the difference between the current reward value and the previous reward value is less than χ 1 , the absolute value of the difference between the current Agent output action and the last Agent output action is less than χ 2 , where χ 1 ≥0, χ 2 ≥0 is the set decision threshold, and finally the driving signal spectrum is determined according to the convergence value of the Agent output action.

[0093] Compared with the prior art, the present invention has the following beneficial effects:

[0094] 1. Based on the conventional phase modulation mathematical relationship and experimental data, the present invention adopts a deep learning algorithm to establish a physical model between the driving electrical signal and the seed source broadening spectrum. Through the customized driving signal design, the precise control of each frequency signal can be achieved, effectively improving the control accuracy and flexibility of the seed source spectrum;

[0095] 2. The present invention can design driving signals quickly and adaptively for complex spectral requirements, effectively compensate for the nonlinear response of the system, and optimize the seed source spectrum;

[0096] 3. The present invention can flexibly and automatically design the drive signal spectrum for different laser systems and application scenarios, optimize the seed source spectrum, and has the characteristics of high efficiency and strong adaptability;

[0097] 4. The present invention can break the spectral design limitations of traditional seed source spectrum broadening, and realize automatic driving signal design and system nonlinear response compensation for different laser systems and target spectra, which can meet application requirements under complex spectral requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] 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:

[0099] Figure 1 A flowchart of a method for optimizing spectrum broadening of a high-power fiber laser seed source according to an embodiment of the present invention;

[0100] Figure 2 This is an experimental structure diagram of a high-power fiber laser seed source spectrum broadening optimization method according to an embodiment of the present invention;

[0101] Figure 3 A schematic diagram of a deep learning algorithm for a high-power fiber laser seed source spectrum broadening optimization method according to an embodiment of the present invention;

[0102] Figure 4 This is a spectrum optimization process diagram of a high-power fiber laser seed source spectrum broadening optimization method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0103] 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.

[0104] Embodiment 1:

[0105] The present invention discloses a method for optimizing the broadening of the spectrum of a high-power fiber laser seed source. The method uses a deep learning algorithm to establish a physical model between a phase-modulated driving electrical signal and a spectrum after phase modulation broadening of a seed source, and uses the experimentally designed driving signal spectrum and the seed source spectrum measurement data as references to establish a neural network. According to the target spectrum of the seed source of a high-power fiber laser in different application scenarios, through multiple iterations, the driving electrical signal and spectrum are reversely designed and optimized, and the spectrum spectrum errors caused by factors including the nonlinear response of the device are compensated. The method can quickly and directly search the driving electrical signal spectrum and realize the broadening of the target spectrum of the seed source.

[0106] According to a high-power fiber laser seed source spectrum broadening optimization method provided by the present invention, Figure 1-Figure 4 As shown, including:

[0107] Step S1: parameter initialization, designing the driving electrical signal according to the target spectrum;

[0108] Specifically, in step S1:

[0109] Initialize the driving electrical signal, agent, experience playback pool, spectral state and define the state transition process. According to the target spectral shape of the seed source, design the driving electrical signal of the corresponding spectral type.

[0110] Initialization objects include Actor network μ θ , Critic Network Q w , Target Actor Network TargetCritic Network and experience replay pool;

[0111] Initialize the Actor network μ with random parameters θ, w θ and Critic Network Q w ;

[0112] Using random parameters Initialize the Target Actor Network and Target Critic Network

[0113] Among them, setting The initial value of is equal to θ, set The initial value is equal to w; the format of experience storage in the experience replay pool is:

[0114] E=(s i ,a i ,r i ,s i+1 ,done)

[0115] Among them, s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed;

[0116] The spectral design of the driving electrical signal includes amplitude, phase, frequency interval, bandwidth and amplitude envelope shape; the spectral information of the seed source includes amplitude, bandwidth, in-band flatness and amplitude envelope shape.

[0117] Step S2: outputting a driving electrical signal to broaden the spectrum of the seed source;

[0118] Specifically, in step S2:

[0119] The signal generator outputs a preset driving electrical signal; the generated driving electrical signal is input to the RF input port of the phase modulator after power amplification, and the seed source outputs a laser with a preset wavelength, which is input to the phase modulator for spectrum modulation and broadening;

[0120] The signal generator includes an arbitrary waveform generator and a digital-to-analog converter;

[0121] The phase modulation spectrum broadening process of the seed source is single-level phase modulation or multi-level phase modulation;

[0122] Step S3: measuring and storing the broadened seed source spectrum information, analyzing the spectrum information to obtain empirical values ​​and storing them;

[0123] In step S3:

[0124] The spectrometer measures and stores the broadened seed source spectrum information;

[0125] Initialize the seed source spectrum state:

[0126] s 0 =(1,I limit (1),0,0,…0)

[0127] The i-th state of the seed source spectrum is defined as:

[0128] s i =(id,I limit (id),I 1 ,I 2 ,…I m )

[0129] Among them, id represents the index of the position to be determined, and the positions are sorted in the order of frequency components from 1 to m. limit (id) represents the maximum power limit of the idth position, I id ,id=1,2,…m represents the power value of the idth frequency component;

[0130] The seed source spectrum state transfer process is defined as:

[0131] From the initial state s 0 =(1,I limit(1),0,0,…0), the agent generates action a 0 =I 1 , set the power value of the frequency component and get reward r 0 =0, update the state to s 1 =(2,I limit (2),I 1 ,0,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0132] From the spectral state s i =(i+1,I limit (i+1),I 1 ,I 2 ,…I i ,…0) starts, Agent generates action a i =I i+1 , set the power value of the i+1th frequency component, and then get the reward r i =0, update the state to s i+1 =(i+2,I limit (i+2),I 1 ,I 2 ,…I i ,I i+1 ,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0133] And so on, until the state is updated to the end state s m =(0,0,I 1 ,I 2 ,…I m ), set the state transfer end flag done = True, indicating that the entire state transfer process is over. When done = True, let r 0 =r 1 =…=r i =…r m =reward, reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed.

[0134] Step S4: training and updating the spectrum optimization network according to the empirical value to obtain the maximum benefit value;

[0135] Specifically, in step S4:

[0136] Train and update the agent based on the experience gained; repeat the training and strategy exploration until the maximum benefit value is obtained;

[0137] The state transfer process E=(si ,a i ,r i ,s i+1 , done) as experience, stored in the experience replay pool, and repeat the state update operation until the specified amount of experience is generated;

[0138] Among them, s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed;

[0139] Minimize the loss according to the following formula:

[0140]

[0141] Loss2=-Q w (s i ,μ θ (s i )),

[0142] Reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed;

[0143] represents the target Q network parameter, and Loss1 represents the Critic network Q w Optimization formula, Loss2 represents the Actor network μ θ Optimization formula, MSE means mean square error;

[0144] Update Critic Network Q w and Actor Network μ θ , Q w (s i ,a i ) represents the critic network Q w Facing state s i and action is a i Output at time, μ θ (s i ) represents the Actor network μ θ Facing state s i Output action, Q w (s i ,μθ (s i )) represents the critic network Q w Facing state s i and action is μ θ (s i ), α is the discount factor.

[0145] Step S5: Establish a corresponding model between the driving electric signal and the seed source spectrum, and optimize the driving electric signal spectrum corresponding to the target spectrum.

[0146] Specifically, in step S5:

[0147] The corrected driving electric signal and the seed source broadened spectrum are analyzed, the corresponding model between the driving electric signal and the seed source spectrum is reconstructed, and the driving electric signal spectrum corresponding to the target spectrum is obtained through an iterative optimization method;

[0148] Repeat the above training update operation, and the training end judgment standard is: the absolute value of the difference between the current reward value and the previous reward value is less than χ 1 , the absolute value of the difference between the current Agent output action and the last Agent output action is less than χ 2 , where χ 1 ≥0, χ 2 ≥0 is the set decision threshold, and finally the driving signal spectrum is determined according to the convergence value of the Agent output action.

[0149] Embodiment 2:

[0150] Embodiment 2 is a preferred example of Embodiment 1, and is used to illustrate the present invention in more detail.

[0151] Those skilled in the art may understand the high-power fiber laser seed source spectrum broadening optimization method provided by the present invention as a specific implementation of a high-power fiber laser seed source spectrum broadening optimization system, that is, the high-power fiber laser seed source spectrum broadening optimization system may be implemented by executing the step flow of the high-power fiber laser seed source spectrum broadening optimization method.

[0152] A high-power fiber laser seed source spectrum broadening optimization system provided by the present invention comprises:

[0153] Module M1: parameter initialization, designing driving electrical signals according to the target spectrum;

[0154] Specifically, in the module M1:

[0155] Initialize the driving electrical signal, agent, experience playback pool, spectral state and define the state transition process. According to the target spectral shape of the seed source, design the driving electrical signal of the corresponding spectral type.

[0156] Initialization objects include Actor network μ θ , Critic Network Q w , Target Actor Network TargetCritic Network and experience replay pool;

[0157] Initialize the Actor network μ with random parameters θ, w θ and Critic Network Q w ;

[0158] Using random parameters Initialize the Target Actor Network and Target Critic Network

[0159] Among them, setting The initial value of is equal to θ, set The initial value is equal to w; the format of experience storage in the experience replay pool is:

[0160] E=(s i ,a i ,r i ,s i+1 ,done)

[0161] Among them, s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed;

[0162] The spectral design of the driving electrical signal includes amplitude, phase, frequency interval, bandwidth and amplitude envelope shape; the spectral information of the seed source includes amplitude, bandwidth, in-band flatness and amplitude envelope shape.

[0163] Module M2: outputs driving electrical signals to broaden the spectrum of the seed source;

[0164] Specifically, in the module M2:

[0165] The signal generator outputs a preset driving electrical signal; the generated driving electrical signal is input to the RF input port of the phase modulator after power amplification, and the seed source outputs a laser with a preset wavelength, which is input to the phase modulator for spectrum modulation and broadening;

[0166] The signal generator includes an arbitrary waveform generator and a digital-to-analog converter;

[0167] The phase modulation spectrum broadening process of the seed source is single-level phase modulation or multi-level phase modulation;

[0168] Module M3: measure and store the broadened seed source spectrum information, analyze the spectrum information to obtain the empirical value and store it;

[0169] In the module M3:

[0170] The spectrometer measures and stores the broadened seed source spectrum information;

[0171] Initialize the seed source spectrum state:

[0172] s 0 =(1,I limit (1),0,0,…0)

[0173] The i-th state of the seed source spectrum is defined as:

[0174] s i =(id,I limit (id),I 1 ,I 2 ,…I m )

[0175] Among them, id represents the index of the position to be determined, and the positions are sorted in the order of frequency components from 1 to m. limit (id) represents the maximum power limit of the idth position, I id ,id=1,2,…m represents the power value of the idth frequency component;

[0176] The seed source spectrum state transfer process is defined as:

[0177] From the initial state s 0 =(1,I limit (1),0,0,…0), the agent generates action a 0 =I 1 , set the power value of the frequency component and get reward r 0 =0, update the state to s 1 =(2,I limit (2),I1 ,0,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0178] From the spectral state s i =(i+1,I limit (i+1),I 1 ,I 2 ,…I i ,…0) starts, Agent generates action a i =I i+1 , set the power value of the i+1th frequency component, and then get the reward r i =0, update the state to s i+1 =(i+2,I limit (i+2),I 1 ,I 2 ,…I i ,I i+1 ,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0179] And so on, until the state is updated to the end state s m =(0,0,I 1 ,I 2 ,…I m ), set the state transfer end flag done = True, indicating that the entire state transfer process is over. When done = True, let r 0 =r 1 =…=r i =…r m =reward, reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed.

[0180] Module M4: Train and update the spectrum optimization network based on empirical values ​​to obtain the maximum benefit value;

[0181] Specifically, in the module M4:

[0182] Train and update the agent based on the experience gained; repeat the training and strategy exploration until the maximum benefit value is obtained;

[0183] The state transfer process E=(s i ,a i ,r i ,s i+1 , done) as experience, stored in the experience replay pool, and repeat the state update operation until the specified amount of experience is generated;

[0184] Among them, si Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed;

[0185] Minimize the loss according to the following formula:

[0186]

[0187] Loss2=-Q w (s i ,μ θ (s i )),

[0188] Reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed;

[0189] represents the target Q network parameter, and Loss1 represents the Critic network Q w Optimization formula, Loss2 represents the Actor network μ θ Optimization formula, MSE means mean square error;

[0190] Update Critic Network Q w and Actor Network μ θ , Q w (s i ,a i ) represents the critic network Q w Facing state s i and action is a i Output at time, μ θ (s i ) represents the Actor network μ θ Facing state s i Output action, Q w (s i ,μ θ (s i )) represents the critic network Q w Facing state s i and action is μ θ (s i) when α is the discount factor. Module M5: Establish the corresponding model between the driving electric signal and the seed source spectrum, and optimize the driving electric signal spectrum corresponding to the target spectrum.

[0191] Specifically, in the module M5:

[0192] The corrected driving electric signal and the seed source broadened spectrum are analyzed, the corresponding model between the driving electric signal and the seed source spectrum is reconstructed, and the driving electric signal spectrum corresponding to the target spectrum is obtained through an iterative optimization method;

[0193] Repeat the above training update operation, and the training end judgment standard is: the absolute value of the difference between the current reward value and the previous reward value is less than χ 1 , the absolute value of the difference between the current Agent output action and the last Agent output action is less than χ 2 , where χ 1 ≥0, χ 2 ≥0 is the set decision threshold, and finally the driving signal spectrum is determined according to the convergence value of the Agent output action.

[0194] Embodiment 3:

[0195] Embodiment 3 is a preferred example of Embodiment 1, and is used to illustrate the present invention in more detail.

[0196] In view of the application requirements, the technical problem to be solved by the present invention is to establish a model for automatically searching for the best driving signal under a given target spectrum, adopt a deep learning algorithm, output actions as different types of driving signal waveform parameters, and rewards are designed as functions related to the spectrum shape, flatness, bandwidth, etc. The present invention can break the spectral design limitations of traditional seed source spectrum broadening, and realize automated driving signal design and system nonlinear response compensation for different laser systems and target spectra, which can meet the application requirements under complex spectral requirements.

[0197] To achieve the above object, the technical solution of the present invention is as follows: a method for optimizing the spectrum broadening of a high-power fiber laser seed source, the method comprising the following steps:

[0198] S1: Initialize the driving electrical signal, the agent, the experience replay pool, the spectral state and define the state transition process. According to the target spectral shape of the seed source, simulate and design the driving electrical signal of the corresponding spectral type.

[0199] S2: The signal generator outputs the driving electrical signal preset in step S1;

[0200] S3: The driving electrical signal generated in step S2 is input to the RF input port of the phase modulator after power amplification, and the seed source outputs a laser with a specific wavelength, which is input to the phase modulator for spectrum modulation and broadening;

[0201] S4: The spectrometer measures and stores the seed source spectrum information after being broadened in step S3;

[0202] S5: parsing the seed source spectrum information obtained in step S4, obtaining corresponding experience values, and storing them;

[0203] S6: Train and update the agent according to the experience value obtained in step S5;

[0204] S7: Repeat training and strategy exploration until the maximum benefit value is obtained;

[0205] S8: Analyze the corrected driving electric signal and the seed source broadened spectrum, reconstruct the corresponding model between the driving electric signal and the seed source spectrum, and obtain the driving electric signal spectrum corresponding to the target spectrum through an iterative optimization method.

[0206] Furthermore, in step S1:

[0207] Initialization objects include Actor network μ θ , Critic Network Q w , Target Actor Network TargetCritic Network and experience replay pool; initialize the Actor network μ with random parameters θ, w θ and Critic Network Q w ; Use random parameters Initialize the Target Actor Network and Target Critic Network Among them, setting The initial value of is equal to θ, set The initial value is equal to w; the format of experience replay pool to store experience is E = (s i ,a i ,r i ,s i+1 , done), where s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the seed source spectrum is updated, and done is the sign of whether the whole state transfer process is completed.

[0208] Furthermore, in step S1:

[0209] The spectral design of the driving electrical signal includes amplitude, phase, frequency interval, bandwidth and amplitude envelope shape; the spectral information of the seed source includes amplitude, bandwidth, in-band flatness and amplitude envelope shape.

[0210] Furthermore, in step S2:

[0211] The signal generator includes an arbitrary waveform generator and a digital-to-analog converter (DAC), which can meet the speed, bandwidth and spectrum design requirements of the driving electrical signal.

[0212] Furthermore, in step S3:

[0213] The phase modulation spectrum broadening process of the seed source can be single-level phase modulation or multi-level phase modulation.

[0214] Furthermore, the step S5 specifically includes the following steps:

[0215] Initialize the seed source spectrum state to s 0 =(1,I limit (1),0,0,…0);

[0216] The i-th state of the seed source spectrum is defined as s i =(id,I limit (id),I 1 ,I 2 ,…I m ), id represents the index of the position to be determined, and the positions are sorted in the order of frequency components from 1 to m. limit (id) represents the maximum power limit of the idth position, I id ,id=1,2,…m represents the power value of the idth frequency component;

[0217] The seed source spectrum state transfer process is defined as:

[0218] From the initial state s 0 =(1,I limit (1),0,0,…0), the agent generates action a 0 =I 1 , set the power value of the frequency component and get reward r 0 =0r 0 =0, update the state to s 1 =(2,I limit (2),I 1,0,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0219] From the spectral state s i =(i+1,I limit (i+1),I 1 ,I 2 ,…I i ,…0) starts, Agent generates action a i =I i+1 , set the power value of the i+1th frequency component, and then get the reward r i =0, update the state to s i+1 =(i+2,I limit (i+2),I 1 ,I 2 ,…I i ,I i+1 ,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0220] And so on, until the state is updated to the end state s m =(0,0,I 1 ,I 2 ,…I m ), set the state transfer end flag done = True, indicating that the entire state transfer process is over. When done = True, let r 0 =r 1 =…=r i =…r m =reward, reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed.

[0221] Further, in step S6:

[0222] The state transfer process E=(s i ,a i ,r i ,s i+1 , done) as experience, stored in the experience replay pool, and repeat the state update operation until the specified amount of experience is generated;

[0223] Minimize the loss according to the following formula,

[0224]

[0225] Loss2=-Q w (s i ,μ θ (s i )),

[0226] Update Critic Network Q w and Actor Network μ θ , Q w (s i ,a i ) represents the critic network Q w Facing state s i and action is a i Output at time, μ θ (s i ) represents the Actor network μ θ Facing state s i Output action, Q w (s i ,μ θ (s i )) represents the critic network Q w Facing state s i and action is μ θ (s i ) when α is the discount factor;

[0227] Furthermore, in step S8:

[0228] Repeat the above training update operation, and the training end judgment standard is: the absolute value of the difference between the current reward value and the previous reward value is less than χ 1 , the absolute value of the difference between the current Agent output action and the last Agent output action is less than χ 2 , where χ 1 ≥0, χ 2 ≥0 is the set decision threshold, and finally the driving signal spectrum is determined according to the convergence value of the Agent output action.

[0229] Embodiment 4:

[0230] Embodiment 4 is a preferred example of Embodiment 1, and is used to illustrate the present invention in more detail.

[0231] like Figure 1 As shown, a high-power fiber laser seed source spectrum broadening optimization method proposed in the present invention comprises the following steps:

[0232] S1: Build a spectrum broadening and measurement experimental platform, such as Figure 2 As shown, initialize the driving electrical signal, agent, experience playback pool, spectral state and define the state transition process:

[0233] Initialization objects include Actor network μ θ , Critic Network Q w, Target Actor Network TargetCritic Network and experience replay pool; initialize the Actor network μ with random parameters θ, w θ and Critic Network Q w ; Use random parameters Initialize the Target Actor Network and Target Critic Network Among them, setting The initial value of is equal to θ, set The initial value is equal to w; the format of the experience replay pool to store experience is E = (s i ,a i ,r i ,s i+1 , done), where s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the seed source spectrum is updated, and done is the sign of whether the whole state transfer process is completed. MATLAB simulation designs the initial driving electrical signal.

[0234] S2: Taking an arbitrary waveform generator as a signal generator as an example, adjust the signal parameters including amplitude, phase, frequency interval, bandwidth and amplitude envelope shape to match the signal parameters preset in step S1, and convert the digital signal into an analog signal output.

[0235] S3: Taking single-stage phase modulation as an example, the output signal of step S2 passes through an RF power amplifier and is input from the RF port of the phase modulator as a modulation driving signal. The seed source is input from the optical port of the phase modulator. After modulation, the spectrum is broadened.

[0236] S4: The spectrometer measures and stores the spectrum information of the seed source after the broadening in step S3, including the amplitude, bandwidth, in-band flatness and amplitude envelope shape;

[0237] S5: Analyze the seed source spectrum information obtained in step S4 to obtain the corresponding empirical value, such as Figure 3 As shown, and stored:

[0238] Initialize the seed source spectrum state to s0 =(1,I limit (1),0,0,…0);

[0239] The i-th state of the seed source spectrum is defined as s i =(id,I limit (id),I 1 ,I 2 ,…I m ), id represents the index of the position to be determined, and the positions are sorted in the order of frequency components from 1 to m. limit (id) represents the maximum power limit of the idth position, I id ,id=1,2,…m represents the power value of the idth frequency component;

[0240] The seed source spectrum state transfer process is defined as:

[0241] From the initial state s 0 =(1,I limit (1),0,0,…0), the agent generates action a 0 =I 1 , set the power value of the frequency component and get reward r 0 =0r 0 =0, update the state to s 1 =(2,I limit (2),I 1 ,0,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0242] From the spectral state s i =(i+1,I limit (i+1),I 1 ,I 2 ,…I i ,…0) starts, Agent generates action a i =I i+1 , set the power value of the i+1th frequency component, and then get the reward r i =0, update the state to s i+1 =(i+2,I limit (i+2),I 1 ,I 2 ,…I i ,I i+1 ,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed;

[0243] And so on, until the state is updated to the end state s m =(0,0,I 1 ,I2 ,…I m ), set the state transfer end flag done = True, indicating that the entire state transfer process is over. When done = True, let r 0 =r 1 =…=r i =…r m =reward, reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed.

[0244] S6: Train and update the agent based on the experience value obtained in step S5:

[0245] The state transfer process E=(s i ,a i ,r i ,s i+1 , done) as experience, stored in the experience replay pool, and repeat the state update operation until the specified amount of experience is generated;

[0246] Minimize the loss according to the following formula,

[0247]

[0248] Loss2=-Q w (s i ,μ θ (s i )),

[0249] Update Critic Network Q w and Actor Network μ θ , Q w (s i ,a i ) represents the critic network Q w Facing state s i and action is a i Output at time, μ θ (s i ) represents the Actor network μ θ Facing state s i Output action, Q w (s i ,μ θ (s i )) represents the critic network Q w Facing state s i and action is μ θ (s i ) when α is the discount factor;

[0250] S7: Repeat training and strategy exploration until the maximum benefit value is obtained;

[0251] S8: Analyze the corrected driving electric signal and the seed source broadened spectrum, reconstruct the corresponding model between the driving electric signal and the seed source spectrum, and obtain the driving electric signal spectrum corresponding to the target spectrum through an iterative optimization method. The training end judgment standard is: the absolute value of the difference between the current reward value and the previous reward value is less than χ 1 , the absolute value of the difference between the current Agent output action and the last Agent output action is less than χ 2 , where χ 1 ≥0, χ 2 ≥0 is the set decision threshold, and finally the driving signal spectrum is determined according to the convergence value of the Agent output action.

[0252] Take the seed source broadened spectrum with wide bandwidth and flat amplitude in the band as an example, Figure 4 As shown in the figure, the agent takes actions according to the spectral state, outputs the driving spectrum parameters, and optimizes its own strategy based on the reward of the measured spectral feedback, and finally determines the best driving signal. The deep deterministic policy gradient algorithm (DDPG) in the deep reinforcement learning algorithm is used as the agent. The policy network plays the role of Actor, which is responsible for external display of output. The Q network is the critic, and the Actor needs to make an action according to the current state of the stage. i , the Q network makes a benefit evaluation on the action output by the Actor at each step, that is, Q w (s,a). The Actor adjusts its strategy based on the judges’ scores and updates the neural network parameters. The Critic rewards r based on the feedback from the spectral measurement environment. i Adjust the scoring strategy and update the Critic's neural network parameter Q w The goal is to maximize the total benefit. The present invention utilizes high-precision spectrum measurement and optimization model, combined with the weak responsivity of a single low-bandwidth phase modulator at high frequencies, to ensure an ideal in-band spectrum while achieving a steep drop in the spectrum, ultimately achieving an optimized design of bandwidth and spectrum.

[0253] Those skilled in the art know that, in addition to implementing the system, device and its various modules provided by the present invention in a purely computer-readable program code, it is entirely possible to implement the same program 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, device and its various modules provided by the present invention can be considered as a hardware component, and the modules included therein for implementing various programs can also be considered as structures within the hardware component; the modules for implementing various functions can also be considered as both software programs for implementing the method and structures within the hardware component.

[0254] 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 optimizing spectrum broadening of a high-power fiber laser seed source, characterized in that: include: Step S1: parameter initialization, designing the driving electrical signal according to the target spectrum; Step S2: outputting a driving electrical signal to broaden the spectrum of the seed source; Step S3: measuring and storing the broadened seed source spectrum information, analyzing the spectrum information to obtain empirical values ​​and storing them; Step S4: training and updating the spectrum optimization network according to the empirical value to obtain the maximum benefit value; Step S5: establishing a corresponding model between the driving electric signal and the seed source spectrum, and optimizing the driving electric signal spectrum corresponding to the target spectrum; In step S3: The spectrometer measures and stores the broadened seed source spectrum information; Initialize the seed source spectrum state: s0=(1,I limit (1),0,0,…0) The i-th state of the seed source spectrum is defined as: s i =(id,I limit (id),I1,I2,…I m ) Among them, id represents the index of the position to be determined, and the positions are sorted in the order of frequency components from 1 to m. limit (id) represents the maximum power limit of the idth position, I id ,id=1,2,…m represents the power value of the idth frequency component; The seed source spectrum state transfer process is defined as: From the initial state s0=(1,I limit (1),0,0,…0), the Agent generates action a0=I1, sets the power value of the frequency component, obtains reward r0=0, and updates the state to s1=(2,I limit (2), I1, 0, ... 0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed; From the spectral state s i =(i+1,I limit (i+1),I1,I2,…I i ,…0) starts, Agent generates action a i =I i+1 , set the power value of the i+1th frequency component, and then get the reward r i =0, update the state to s i+1 =(i+2,I limit (i+2),I1,I2,…I i ,I i+1 ,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed; And so on, until the state is updated to the end state s m =(0,0,I1,I2,…I m ), set the state transfer end flag done = True, indicating that the entire state transfer process is completed. When done = True, set r0 = r1 = ... = r i =…r m =reward, reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed; In step S4: Train and update the agent based on the experience gained; repeat the training and strategy exploration until the maximum benefit value is obtained; The state transfer process E=(s i ,a i ,r i ,s i+1 , done) as experience, stored in the experience replay pool, and repeat the state update operation until the specified amount of experience is generated; Among them, s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed; Minimize the loss according to the following formula: Loss2=-Q w (s i ,m θ (s i )), Reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed; represents the target Q network parameter, and Loss1 represents the Critic network Q w Optimization formula, Loss2 represents the Actor network μ θ Optimization formula, MSE means mean square error; Update Critic Network Q w and Actor Network μ θ , Q w (s i ,a i ) represents the critic network Q w Facing state s i and action is a i Output at time, μ θ (s i ) represents the Actor network μ θ Facing state s i Output action, Q w (s i ,μ θ (s i )) represents the critic network Q w Facing state s i and action is μ θ (s i ), α is the discount factor.

2. The high-power fiber laser seed source spectrum broadening optimization method according to claim 1, characterized in that: In step S1: Initialize the driving electrical signal, agent, experience playback pool, spectral state and define the state transition process. According to the target spectral shape of the seed source, design the driving electrical signal of the corresponding spectral type. Initialization objects include Actor network μ θ , Critic Network Q w , Target Actor Network Target Critic Network and experience replay pool; Initialize the Actor network μ with random parameters θ, w θ and Critic Network Q w ; Using random parameters Initialize the Target Actor Network and Target Critic Network Among them, setting The initial value of is equal to θ, set The initial value is equal to w; the format of experience storage in the experience replay pool is: E=(s i ,a i ,r i ,s i+1 ,done) Among them, s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed; The spectral design of the driving electrical signal includes amplitude, phase, frequency interval, bandwidth and amplitude envelope shape; the spectral information of the seed source includes amplitude, bandwidth, in-band flatness and amplitude envelope shape.

3. The high power fiber laser seed source spectrum broadening optimization method according to claim 1, characterized in that: In step S2: The signal generator outputs a preset driving electrical signal; The generated driving electrical signal is input into the RF input port of the phase modulator after power amplification, and the seed source outputs laser light of a preset wavelength, which is input into the phase modulator for spectrum modulation and broadening; The signal generator includes an arbitrary waveform generator and a digital-to-analog converter; The phase modulation spectrum broadening process of the seed source is single-level phase modulation or multi-level phase modulation.

4. The high-power fiber laser seed source spectrum broadening optimization method according to claim 1, characterized in that: In step S5: The corrected driving electric signal and the seed source broadened spectrum are analyzed, the corresponding model between the driving electric signal and the seed source spectrum is reconstructed, and the driving electric signal spectrum corresponding to the target spectrum is obtained through an iterative optimization method; Repeat the training update operation. The training end judgment criteria are: the absolute value of the difference between the current reward value and the previous reward value is less than χ1, and the absolute value of the difference between the current Agent output action and the previous Agent output action is less than χ2, where χ1≥0 and χ2≥0 are the set judgment thresholds. Finally, the driving signal spectrum is determined according to the convergence value of the Agent output action.

5. A high-power fiber laser seed source spectrum broadening optimization system, characterized in that: include: Module M1: parameter initialization, designing driving electrical signals according to the target spectrum; Module M2: outputs driving electrical signals to broaden the spectrum of the seed source; Module M3: measure and store the broadened seed source spectrum information, analyze the spectrum information to obtain the empirical value and store it; Module M4: Train and update the spectrum optimization network based on empirical values ​​to obtain the maximum benefit value; Module M5: Establish a corresponding model between the driving electric signal and the seed source spectrum, and optimize the driving electric signal spectrum corresponding to the target spectrum; In the module M3: The spectrometer measures and stores the broadened seed source spectrum information; Initialize the seed source spectrum state: s0=(1,I limit (1),0,0,…0) The i-th state of the seed source spectrum is defined as: s i =(id,I limit (id),I1,I2,…I m ) Among them, id represents the index of the position to be determined, and the positions are sorted in the order of frequency components from 1 to m. limit (id) represents the maximum power limit of the idth position, I id ,id=1,2,…m represents the power value of the idth frequency component; The seed source spectrum state transfer process is defined as: From the initial state s0=(1,I limit (1),0,0,…0), the Agent generates action a0=I1, sets the power value of the frequency component, obtains reward r0=0, and updates the state to s1=(2,I limit (2), I1, 0, ... 0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed; From the spectral state s i =(i+1,I limit (i+1),I1,I2,…I i ,…0) starts, Agent generates action a i =I i+1 , set the power value of the i+1th frequency component, and then get the reward r i =0, update the state to s i+1 =(i+2,I limit (i+2),I1,I2,…I i ,I i+1 ,…0), set the state transfer end flag done = False, indicating that the entire state transfer process is not completed; And so on, until the state is updated to the end state s m =(0,0,I1,I2,…I m ), set the state transfer end flag done = True, indicating that the entire state transfer process is completed. When done = True, set r0 = r1 = ... = r i =…r m =reward, reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed; In the module M4: Train and update the agent based on the experience gained; repeat the training and strategy exploration until the maximum benefit value is obtained; The state transfer process E=(s i ,a i ,r i ,s i+1 , done) as experience, stored in the experience replay pool, and repeat the state update operation until the specified amount of experience is generated; Among them, s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed; Minimize the loss according to the following formula: Loss1=MSE[Q w (s i ,a i ),reward+αQ w (s',a')] Loss2=-Q w (s i ,m θ (s i )), Reward represents the reward value calculated based on the current spectrum type after the entire state transfer process is completed; Q w (s',a') represents the target Q network parameters, Loss1 represents the Critic network Q w Optimization formula, Loss2 represents the Actor network μ θ Optimization formula, MSE means mean square error; Update Critic Network Q w and Actor Network μ θ , Q w (s i ,a i ) represents the critic network Q w Facing state s i and action is a i Output at time, μ θ (s i ) represents the Actor network μ θ Facing state s i Output action, Q w (s i ,μ θ (s i )) represents the critic network Q w Facing state s i and action is μ θ (s i ), α is the discount factor.

6. The high-power fiber laser seed source spectrum broadening optimization system according to claim 5, characterized in that: In the module M1: Initialize the driving electrical signal, agent, experience playback pool, spectral state and define the state transition process. According to the target spectral shape of the seed source, design the driving electrical signal of the corresponding spectral type. Initialization objects include Actor network μ θ , Critic Network Q w , Target Actor Network μ θ , Target Critic Network Q w and experience replay pool; Initialize the Actor network μ with random parameters θ, w θ and Critic Network Q w ; Initialize the Target Actor network μ with random parameters θ, w θ and Target Critic Network Q w ; The initial value of θ is set equal to θ, and the initial value of w is set equal to w. The format of experience storage in the experience replay pool is: E=(s i ,a i ,r i ,s i+1 ,done) Among them, s i Indicates the current seed source spectrum state, a i Indicates that the Agent is based on the current state s i The action generated, that is, the spectrum of the driving electrical signal, r i Indicates that the Agent is facing state s i Take action a i Rewards received, i+1 Indicates that the Agent takes action a i After that, the state of the seed source spectrum after update, done is the sign of whether the whole state transfer process is completed; The spectral design of the driving electrical signal includes amplitude, phase, frequency interval, bandwidth and amplitude envelope shape; the spectral information of the seed source includes amplitude, bandwidth, in-band flatness and amplitude envelope shape.

7. The high-power fiber laser seed source spectrum broadening optimization system according to claim 5, characterized in that: In the module M2: The signal generator outputs a preset driving electrical signal; The generated driving electrical signal is input into the RF input port of the phase modulator after power amplification, and the seed source outputs laser light of a preset wavelength, which is input into the phase modulator for spectrum modulation and broadening; The signal generator includes an arbitrary waveform generator and a digital-to-analog converter; The phase modulation spectrum broadening process of the seed source is single-level phase modulation or multi-level phase modulation.

8. The high-power fiber laser seed source spectrum broadening optimization system according to claim 5, characterized in that: In the module M5: The corrected driving electric signal and the seed source broadened spectrum are analyzed, the corresponding model between the driving electric signal and the seed source spectrum is reconstructed, and the driving electric signal spectrum corresponding to the target spectrum is obtained through an iterative optimization method; Repeat the training update operation. The training end judgment criteria are: the absolute value of the difference between the current reward value and the previous reward value is less than χ1, and the absolute value of the difference between the current Agent output action and the previous Agent output action is less than χ2, where χ1≥0 and χ2≥0 are the set judgment thresholds. Finally, the driving signal spectrum is determined according to the convergence value of the Agent output action.

Images