Laser output power stabilizing control method based on HLOA-SVR-AIMC

Abstract

The present application relates to the technical field of laser output power stabilization, and particularly discloses a laser output power stabilization control method based on HLOA-SVR-AIMC, which is based on the physical mechanism of Bragg diffraction, establishes a nonlinear mapping relationship between diffraction efficiency and driving voltage, uses a regression support vector machine to establish an internal model and an inverse model of the system, takes the inverse model as a controller of internal model control, and introduces an adaptive feedback filter to reduce the sensitivity of internal model control to modeling errors and improve the robustness of the system. The method and results proposed by the present application can be widely applied to signal system identification, control and signal processing, etc., expand the scope of application of the angle lizard optimization algorithm, the regression support vector machine and the adaptive internal model control in the engineering technical field, and make the system more worthy of promotion and use in trace gas concentration detection.

Description

Technical Field

[0001] This invention relates to a method for stabilizing and controlling optical output power, specifically a method for stabilizing and controlling laser output power based on HLOA-SVR-AIMC, belonging to the field of laser output power stabilization technology. Background Technology

[0002] Since the world’s first ruby ​​solid-state pulsed laser was invented in the 1960s, laser development has entered a fast track and has quickly intersected and permeated with multiple disciplines to form many new research fields. In recent decades, the application of lasers has gradually developed towards automation and intelligence. Laser control has received more and more attention, and stable modulation of laser output power plays an indispensable role in the field of laser control. It is a popular research topic for obtaining high-precision and high-stability laser beams.

[0003] With the development of feedback control theory and technology, external cavity modulation for laser output power stabilization has become mainstream, and laser output power stabilization based on acousto-optic modulators has become the most widely used external cavity modulation path at present. However, observations from existing literature reveal that most laser power stabilization control based on acousto-optic modulation and using feedback technology utilizes proportional-integral-differential (PID) control or its improved versions. While PID control is often effective for linear systems, it generally struggles to achieve satisfactory control performance when applied to nonlinear systems. Although a few researchers have achieved laser output power stabilization using PID control, this is based on a series of simplifications to the laser output power stabilization system (e.g., neglecting the strong nonlinear relationship between the diffraction efficiency of the acousto-optic modulator and the driving power of the driver). Therefore, the designed controllers still suffer from robustness risks and insufficient adaptability in practical applications.

[0004] Therefore, there is an urgent need for a simple, convenient, fast, effective, stable and reliable extracavity modulation strategy and control method for stabilizing laser output power, especially one that can still work normally and maintain good robustness when the laser output power stabilization system is disturbed. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to achieve long-term stability of laser output power in the existing laser technology field. It provides a laser output power stabilization system and control method based on HLOA-SVR-AIMC. This method is also applicable to maintaining stable and reliable laser output power when the system is disturbed.

[0006] The control method of this invention differs from existing laser output power stabilization methods and approaches. It mainly consists of three parts: horned lizard optimization algorithm (HLOA), support vector regression (SVR), and adaptive internal model control (AIMC). This enables the laser to be stabilized quickly, effectively, and for a long time at any given laser power (or voltage), with relative fluctuations not exceeding 0.12%. At the same time, it can maintain stable and reliable performance even when the laser output power is disturbed during the stabilization process. This invention first establishes a nonlinear mapping relationship between diffraction efficiency and driving voltage based on the physical mechanism of Bragg diffraction. It then uses a regressive support vector machine (SVM) to build the system's internal and inverse models, using the inverse model as the controller for internal model control. An adaptive feedback filter is introduced to reduce the sensitivity of internal model control to modeling errors, improving the system's robustness. Secondly, for the parameters of the regressive SVM (e.g., penalty factor, kernel width, and insensitive loss coefficient), a horn-lizard optimization algorithm with a multi-strategy optimization mechanism is introduced to perform a global adaptive search for the parameters. To avoid model overfitting and physical distortion problems caused by traditional single-error optimization, a multi-objective fitness function is designed that couples prediction accuracy, total system energy consumption, model complexity, and physical monotonicity constraints. Finally, to achieve precise control of the laser output power stabilization system, the advantages of the horn-lizard optimization algorithm, regressive SVM, and adaptive internal model control in handling parameter uncertainty, strong nonlinearity, and susceptibility to disturbances are combined, and a new approach is proposed. An adaptive internal model control strategy based on regression-based support vector machine and horned lizard optimization algorithm, namely the HLOA-SVR-AIMC control method, is proposed. Finally, the feasibility and effectiveness of this control method are verified through computer simulation, and a comparative analysis with other control methods demonstrates its superiority. Furthermore, an experimental platform further confirms that when handling a laser output power stabilization system, the laser beam output power fluctuation is compressed from ±0.95% to ±0.08% without delay from the free-running state to the closed-loop control state, meeting the requirements of high-precision spectral measurement and providing new ideas and references for the stabilization control research of other precision optoelectronic instruments. To address the problems in the existing technology, this invention can be achieved through the following technical solutions: The laser output power stabilization control method based on HLOA-SVR-AIMC (1) In order to design an adaptive feedback filter to reduce the sensitivity of internal model control to modeling error, before entering the laser output power stabilization controller design, the nonlinear mapping relationship between diffraction efficiency and driving voltage is established using the physical mechanism of Bragg diffraction. The expression is as follows: , in Indicates the efficiency of diffraction light. Indicates the wavelength of the incident light. The gain coefficient for the electro-acoustic power conversion of the acousto-optic driver. Indicates the interaction length. Indicates the height of the transducer electrode. Indicates the driving voltage. This represents the acousto-optic quality factor of the crystal.

[0007] (2) Based on the regression-type support vector machine, the internal model and inverse model of the laser output power stabilization system are established. The inverse model is used as the controller for controlling the laser output power stabilization system. Combined with step (1), an adaptive feedback filter is introduced. Therefore, the following designs are made respectively: Internal model of the laser output power stabilization system: Assuming the input and output of the entire system can be expressed as , Where y(k) and u(k) are the output and input of the laser output power stabilization system, respectively. It is a nonlinear function. For output delay, For input delay, This is the total delay of the system output relative to the input, and generally... To determine the formula The specific form and related parameter values ​​require the acquisition of N sets of input / output datasets from the laser output power stabilization controlled system. Based on this data, a regression support vector machine is used to establish the internal model of the laser output power stabilization system, and a training sample set D1 is constructed. , in The input characteristics are (i.e., the timing of the laser output and control input of the system over a period of time). For output labels (i.e., hysteresis) (the actual output power of the laser in the step), and The dimension is ,in The dimension is After determining the training sample set D1, a regression-type support vector machine is used to train the samples and establish an internal model of the laser output power stable controlled system.

[0008] Inverse model of the laser output power stabilization system: Similar to the way the internal model is established, it is assumed that... If it is reversible, then the inverse model of the laser output power stabilization system is: , In the formula This is a pending item, and its specific form and related parameters are similar to those of the internal model, i.e., data collection. Input / output dataset of a laser output power stabilization controlled system Based on the collected data, a regression support vector machine is used to establish the inverse model of the laser output power stabilization system, and a training sample set D2 is constructed. , in Input features for the inverse model, Output labels (i.e., the control variables to be predicted) for the inverse model, and The dimension is ,in The dimension is After determining the training sample set D2, a regression-type support vector machine is used to train the samples and establish the inverse model of the laser output power stabilization system. This inverse model is to be used as the controller for the internal model control.

[0009] Adaptive feedback filter for laser output power stabilization system: Design an adaptive filter in the feedback channel of the controlled system. , Where adaptive filtering parameters , , It is the driving voltage. It is a complex variable. This is a comprehensive coefficient related to the acousto-optic crystal material and optical path structure. Based on the basic filter constant, This is the gain sensitivity adjustment coefficient. This represents the system's maximum local gain.

[0010] (3) Based on step (2), first set the reference input power signal (or voltage signal), and enter the comparator together with the feedback channel signal. Then connect it to the controller (i.e., the inverse model), and then enter the module after the controlled object and the internal model are connected in parallel to form the output power signal (or voltage signal) and the feedback signal. At the same time, the feedback signal enters the adaptive feedback filter and then enters the comparator to form a closed loop, and then obtains the structure diagram of the laser output power stable controlled system.

[0011] (4) Based on EQ \o\ac(○,1) and EQ \o\ac(○,2) in step (2), and combined with the system structure diagram, the parameters (i.e., penalty factor, kernel width, and insensitive loss coefficient) of the regression support vector machine are globally adaptively searched using the horned lizard optimization algorithm, where the regression function can be expressed as , In the formula , It is a pair of Lagrange multipliers, and Only a small number of support vectors are retained at that time. It is the bias constant. It is the kernel width. For support vectors, This represents the number of support vectors. Applying the regression function... To approximate the training sample set , ;when At that time, here The datasets are transformed into the datasets of EQ \o\ac(○,1) and EQ \o\ac(○,2) in step (2). Meanwhile, the training sample set (or dataset) satisfies the constraints. , in These represent the lower and upper slack variables, respectively. This represents the insensitive loss coefficient.

[0012] The optimization goal is , in Indicates the confidence range. This represents the penalty factor.

[0013] (5) Based on EQ \o\ac(○,3) in step (2) and step (4), the fitness function is considered by integrating the three aspects of prediction accuracy, model complexity and physical logic consistency in a weighted manner, so as to overcome the model overfitting and physical distortion problems caused by single error optimization in the control process. That is, the ultrasonic power output by the acoustic-optical driver is considered in the actual circuit. With the input modulation voltage The relationship between them: ,in This is the power gain coefficient of the driver. At this point, the total energy consumption of the laser output power stabilization system is... It can be represented as: , Where Δt represents the sampling period (or sampling time).

[0014] Combining the prediction accuracy, model complexity, total energy consumption, and physical model consistency of regression-based support vector machines, a multi-objective fitness function coupled with physical mechanism constraints is constructed as follows: , RMSE is the root mean square error of the training sample set (or dataset), ensuring that the model optimized by the horned lizard optimization algorithm has extremely high fitting accuracy in forward prediction and inverse control. The ratio of support vectors to the total number of samples is used to reduce model complexity and improve the execution efficiency of the controller. This is a physical consistency penalty term. If the parameters generated by the horned lizard optimization algorithm cause the model to experience severe oscillations that are not physically fluctuating within the full range of operating voltage (0.2~0.9 V), this term will impose a very large penalty value; E is the total energy consumption. These represent the weights of each item.

[0015] (6) To further optimize the design, five sets of sample data (or datasets) from different time periods were sampled to optimize the parameters in the control process. The system's performance in building the internal model and inverse model was compared with that of the traditional regression support vector machine and the regression support vector machine based on particle swarm optimization. The root mean square error (RMSE) and the coefficient of determination (R²) were used to evaluate the system's performance. 2 Quantize the predicted signal-to-noise ratio (SNR).

[0016] (7) Based on the above steps, the feasibility and effectiveness of the control method were verified by computer simulation and compared with other control methods, demonstrating the superiority of the proposed control method. At the same time, by building an experimental platform, it was further confirmed that when dealing with the laser output power stabilization system, the laser beam output power fluctuation was compressed from ±0.95% to ±0.08% without delay from the free running state to the closed-loop control state, meeting the requirements of high-precision spectral measurement and providing new ideas and references for the stabilization control research of other precision optoelectronic instruments.

[0017] The beneficial effects of this invention are: This invention, based on the HLOA-SVR-AIMC laser output power stabilization control method, demonstrates advantages in handling objects with uncertain parameters, strong nonlinearity, and susceptibility to disturbances. It addresses the robustness risks and insufficient adaptability issues of laser output power stabilization systems in practical applications. The invention utilizes a regression-type support vector machine to establish the system's internal and inverse models, using the inverse model as the controller for internal model control. An adaptive feedback filter is introduced to reduce the sensitivity of internal model control to modeling errors, improving system robustness. Furthermore, a lizard optimization algorithm is applied to the parameters of the regression-type support vector machine for global adaptive search, avoiding model overfitting and physical distortion problems caused by traditional single-error optimization. Simultaneously, a multi-objective fitness function is designed, coupling prediction accuracy, total system energy consumption, model complexity, and physical monotonicity constraints, to achieve adaptive and robust control of the laser output power stabilization system. In practical applications, the laser output power stabilization control method of this application enables the controller to achieve better tracking performance for any set reference input power signal (or voltage signal). The method and results proposed in this invention can be widely applied to many fields such as signal system identification, control and signal processing, expanding the scope of the horned lizard optimization algorithm, regression support vector machine and adaptive internal model control in the field of engineering technology, making the system more worthy of promotion and use in the detection of trace gas concentrations. Attached Figure Description

[0018] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0019] Figure 1 This is a structural diagram of the control process of the method of the present invention.

[0020] Figure 2 In this embodiment of the invention, the output and error graph of the internal model are established using a regression-type support vector machine.

[0021] Figure 3 This is an embodiment of the invention, verifying the output and error graph of the established internal model.

[0022] Figure 4 In this embodiment of the invention, the output and error graph of the inverse model are established using a regression-type support vector machine.

[0023] Figure 5 This is an embodiment of the invention that verifies the output and error graph of the established inverse model.

[0024] Figure 6 In this embodiment of the invention, the tracking effect diagram and error diagram are obtained by using internal model control based on a regression support vector machine to make the system track a square wave signal.

[0025] Figure 7These are tracking effect diagrams and error diagrams obtained in this embodiment of the invention when the system tracks a square wave signal by adding a disturbance signal using internal model control based on a regression support vector machine.

[0026] Figure 8 This is an experimental model for controlling the laser output power stabilization system in this embodiment of the invention.

[0027] Figure 9 This is the control platform for the laser output power stabilization system built in this embodiment of the invention.

[0028] Figure 10 This is a comparison diagram of the internal model identification error of the laser output power stabilization control system in an embodiment of the present invention.

[0029] Figure 11 This is a comparison diagram of the inverse model identification error of the laser output power stabilization control system in an embodiment of the present invention.

[0030] Figure 12 This is the HOLA fitness change curve in an embodiment of the present invention.

[0031] Figure 13 This is a test of the long-term stability performance of the laser output power stabilization control system in this embodiment of the invention.

[0032] Figure 14 This is an anti-interference test of the laser output power stabilization control system in this embodiment of the invention.

[0033] Figure 15 This is the adaptive parameter variation curve of the laser output power stabilization control system in this embodiment of the invention.

[0034] Figure 16 This is a comparison of the simulation effects of three control methods in an embodiment of the present invention.

[0035] Figure 17 This is a comparison of simulation errors of three control methods in an embodiment of the present invention. Detailed Implementation

[0036] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Please see Figure 1-17 A laser output power stabilization control method based on HLOA-SVR-AIMC, such as Figure 1 As shown, where This indicates the reference input power signal (or voltage signal) to be set. The sampling switch operates according to a period of The discrete signal formed by sampling the reference input power signal (or voltage signal) to be set. Indicates the error signal. The laser output power signal (or voltage signal) is generated by... The signal is recovered from the discrete signal by passing through a zero-order hold (ZOH). This indicates that the laser output power stabilization system is the controlled object, while... These are filters, all of which are obtained through Laplace transform. and These represent the internal model and inverse model built using a regression-based support vector machine, respectively, while the inverse model... They are also controllers used for internal model control; they are all made by The transformation yields, and These represent the outputs of the internal model and the inverse model, respectively. It is an external interference in the control process of the laser output power stabilization system. It is its discrete form. Figure 1 In This represents a multi-head time-delay system. Therefore, the output of the laser output power stabilization control can be expressed as: , in It is a complex variable. and Yes and The results were obtained by performing Laplace transforms on each of them. and Also for and Perform separately Obtained through inverse transformation. At this point, the internal model established by the regression-type support vector machine... If infinitely close (Ideal situation is) And the constructed inverse model (or controller) satisfies When the reference input power signal (or voltage signal) is given, the designed control system can track the input without steady-state error.

[0038] The current goal is to use regression support vector machines to build the internal and inverse models of the control system, based on... Figure 1 The structural diagram shown assumes... and These are the internal model function and the inverse model (or controller) function obtained respectively based on training on the dataset. It should be noted that: if here... and If they are linear functions, then their Laplace transform (or Inverse transform) are respectively (or )and (or However, in general, they are all nonlinear functions and are not suitable for use. (or )and (or However, Figure 1 Still used in China (or )and (or This statement is merely to illustrate the stability of the control system (or its zero-steady-error tracking performance) and does not specify... and It must be a linear function. Furthermore, This represents the predicted output of the internal model. It is the predicted output of the inverse model (or controller). Therefore, The expression can be written as: , in It may not be a linear function; it could be a nonlinear function.

[0039] like Then it is: .

[0040] same, Then it is: , in It may not be a linear function; it could be a nonlinear function.

[0041] like Then it is: .

[0042] The actual input to the inverse model (or controller) is: , in The reference input power signal (or voltage signal) to be set. This is the actual output of the laser output power stabilization system.

[0043] The following steps will be taken to first construct the internal model of the laser output power stabilization system: (1) Determine the sampling time Initialize the system and sample the system's input values ​​at each sampling time. and output value ,in This represents the total number of simulations. In this step, data preprocessing is required, and multiple sets of samples (at least 10 sets) need to be taken, with each set having the same dimension, to facilitate subsequent error verification to ensure compliance with requirements.

[0044] (2) Determine the system input delay and output delay and the total delay of the output relative to the input. Construct a training sample set (or dataset) for building the regression-type support vector machine used to build the internal model: .

[0045] (3) Using the horned lizard optimization algorithm, the parameters of the regression support vector machine are determined, including: kernel function type and its parameters, regularization parameters, loss function and its parameters, etc.

[0046] (4) Apply the training sample set (or dataset) , in which settings ,in accordance with Figure 1 The internal model part is used to train a regression-type support vector machine and build the internal model.

[0047] (5) After training is completed, calculate the training error. If the error does not meet the requirements, combine with step (1) to change the training sample set (or dataset) and repeat the process. Figure 1 The internal model is built by training a regression-type support vector machine in the internal model part until the training error meets the requirements.

[0048] (6) Then select the system input / output values ​​respectively (i.e. and ) Several other typical signals (e.g., random signals, step signals, square wave signals, or sine signals) are used as verification signals to acquire the output value of the acquisition system. Obtain the training sample set (or dataset). Train a regression support vector machine to build an internal model. After training, calculate the training error to test the internal model. If the test does not meet the requirements, return to step (1) to resample the training sample set (or dataset) and continue training and building the internal model until the requirements are met. At this point, the internal model building stage ends, and the internal model function is obtained. .

[0049] Next, we construct the inverse model of the laser output power stabilization system. This part is similar to the internal model establishment process described above, and the specific steps are as follows: (7) Determine the sampling time to be the same as the internal model building process. Initialize the system and sample the system's input values ​​at each sampling time. and output value ,in This represents the total number of simulations. Data preprocessing is essential in this step, and multiple sets of samples (at least 10 sets) must be taken, with each set having the same dimension, to facilitate subsequent error verification to ensure compliance with requirements.

[0050] (8) Determine the input delay of the system, which is the same as the internal model building process. and output delay and the total delay between the output and the input. Construct a training sample set (or dataset) for building a regression-type support vector machine to establish the inverse model: .

[0051] (9) Use the horned lizard optimization algorithm to determine the parameters of the regression support vector machine, including: kernel function type and its parameters, regularization parameters, loss function and its parameters, etc.

[0052] (10) Apply the training sample set (or dataset) , in which settings ,in accordance with Figure 1 The structure of the inverse model part is used to train a regression-type support vector machine and build an inverse model.

[0053] (11) After training is completed, calculate the training error. If the error does not meet the requirements, combine with step (7) to change the training sample set (or dataset) and repeat the process. Figure 1 The inverse model is constructed by training a regression-type support vector machine in the middle of the model until the training error meets the requirements.

[0054] (12) Then select the system input / output values ​​respectively (i.e. and ) Several other typical signals (e.g., random signals, step signals, square wave signals, or sine signals) are used as verification signals to acquire the output value of the acquisition system. Obtain the training sample set (or dataset). Train a regression support vector machine, build an inverse model, and calculate the training error to test the inverse model. If the test fails to meet the requirements, return to step (7) to resample the training sample set (or dataset) and continue training and building the inverse model until the requirements are met. At this point, the inverse model building stage ends, and the inverse model function is obtained. .

[0055] Finally, the internal mode control of the laser output power stabilization system is performed, and the specific steps are as follows: (13) Select a reference input power signal (or voltage signal) and initialize the system's input / output.

[0056] (14) At each sampling time ,when At that time, the actual input for constructing the inverse model (controller) is: .

[0057] Then according to the formula: Calculate the actual output of the inverse model (controller) .

[0058] (15) At each sampling time ,when At that time, the actual input for constructing the internal model: .

[0059] Then according to the formula: Calculate the actual output of the internal model .

[0060] (16) Based on the formula: Calculate the output of the controlled object .

[0061] (17) According to the formula: ,calculate .

[0062] (18) If If it does, then exit the loop; otherwise... Return to step (14) and iterate again.

[0063] (19) Calculate the control error If the control error does not meet the requirements, return to step (13) and reselect the reference input power signal (or voltage signal), or return to step (1) to retrain the system, establish the internal model and inverse model, and implement internal model control until the requirements are met. It should be noted that if the reference input power signal (or voltage signal) is not selected properly, it will also lead to poor control performance.

[0064] (20) If the control error and effect meet the requirements, the relevant results (such as data, images and tables) are output, and the entire laser output power stabilization system control process ends.

[0065] The effectiveness of the method of the present invention is illustrated below through an example: If the nonlinear model is taken as... .

[0066] Establish an internal model: when taking the sampling period Total number of simulations The input is a random signal with an amplitude of 2. The kernel function type is chosen as the radial basis function (RBF) kernel function, and the loss function is chosen as... The insensitivity loss function is determined using the horned lizard optimization algorithm, which sets the kernel width parameter, regularization parameter, and insensitivity parameter of the loss function. The internal model output, built using a regression-based support vector machine, is as follows: Figure 2 As shown in (a), the modeling error is as follows: Figure 2 As shown in (b). From Figure 2 It can be seen that the internal model built using a regression-based support vector machine has high accuracy. Now, we will use a square wave signal as the input to validate the established internal model. The kernel function type will remain the RBF kernel function, and the loss function will also remain [missing information]. The insensitive loss function is determined using the horned lizard optimization algorithm, which also determines the kernel width parameter, regularization parameter, and insensitivity parameter of the loss function. The output of the established internal model is then validated. Figure 3 As shown in (a), the modeling error is as follows: Figure 3 As shown in (b). From Figure 3 It can be seen that the modeling error reaches Order of magnitude.

[0067] The inverse model is established as follows: the sampling period, total number of simulations, input signal, kernel function, and loss function are selected in the same way as the internal model, and the relevant parameters still use the horned lizard optimization algorithm. The output of the inverse model established using a regression-based support vector machine is as follows. Figure 4 As shown in (a), the modeling error is as follows: Figure 4 As shown in (b). To verify the established inverse model, a sinusoidal signal was chosen as the input, resulting in the following: Figure 5 (a) and Figure 5 (b) shows the inverse model output and error. From Figure 4 and Figure 5 It can be seen that the modeling error reaches The inverse model, which is on the order of magnitude larger, also has high accuracy.

[0068] Internal model control: Utilizing the previously established internal model and inverse model, and combining them with the original system to form a system such as... Figure 1 The system shown is now subjected to internal model control. First, a square wave signal is used as the input signal to verify the tracking effect after implementing internal model control. Then, in... and Noise or external disturbances with amplitudes of 0.6 and -0.6 are applied to the system to test its robustness under internal model control. Finally, simulation results are obtained as follows: Figure 6 and Figure 7 The tracking results and error graphs are shown. Figure 6 and Figure 7 It can be seen that the system has strong tracking accuracy and anti-interference ability for square wave signals under internal model control.

[0069] The above example illustrates that the proposed HLOA-SVR-AIMC can identify and control nonlinear models. Next, we will apply HLOA-SVR-AIMC to control a laser output power stabilization system, such as... Figure 8 and Figure 9 As shown, where Figure 9 According to Figure 8 The experimental platform mainly consists of the following components: a laser controller (or driver), a distributed feedback laser (DFB laser), an acousto-optic modulator (AOM), a polarizer, a collimator, a photodetector (PD), a data acquisition card (DAQ card), an acousto-optic driver, and some optical mirrors. The specific control flow is as follows: (1) In the optical path transmission section, the DFB laser emits an initial laser beam under the control of the driver. First, the beam is shaped by a collimator, and then the polarization state of the beam is modulated by a polarizer so that the polarization direction of the incident light matches the wave vector of the acoustic wave in the AOM, thereby achieving the initial diffraction efficiency. The modulated beam is spatially filtered by a grating to remove the 0th order diffracted light and stray light, and then guided by a mirror to finally converge to the PD.

[0070] (2) In the closed-loop feedback section, the PD converts the detected real-time optical power into a corresponding voltage signal, which is then acquired by the DAQ card and uploaded to the host computer. The control algorithm inside the host computer processes this real-time signal, calculates the required control compensation, and commands the DAQ card to output the corresponding analog voltage to the acousto-optic driver. This control voltage adjusts the RF output of the acousto-optic driver in real time, directly changing the instantaneous diffraction efficiency of the AOM, thereby forming a complete control closed loop, and finally locking the laser power output at the preset target value.

[0071] Next, we will first perform data preprocessing. Because AOM is highly sensitive to outliers, if there are significant noise-induced abrupt changes in the training data, the internal model and inverse model may learn incorrect control commands. Therefore, we incorporate the quadrature method for data cleaning. In this invention, we assume that the original output dataset of the internal model / inverse model is... The corresponding original input dataset is Define quartiles and interquartile range (IQR), where To sort all the data from smallest to largest, the number in the 25th position means that 25% of the data are smaller than it. This refers to the number ranked in the 75th percentile, meaning that 75% of the data are smaller than it. The interquartile range (IQR) is... and The difference is then used to define the normal numerical range. Output within this range and input It was preserved, among which is a positive integer.

[0072] Furthermore, due to the differences between the input and output data, although the magnitude difference is not large, in high-dimensional space, the radial basis function of SVR calculates the Euclidean distance. If the numerical range of a certain feature is very large, it will dominate the distance calculation, causing features with small values ​​to be ignored. After normalization, the features of all dimensions are on the same order of magnitude, ensuring that the contribution of each input feature to the model is fair. That is, the normalization expression is: ; in For the raw data Normalized data, The minimum value in the original data. The maximum value in the original data is used as the base value, and the same normalization process is applied to the output data. Finally, the output and input data of the model prediction results are denormalized, and after denormalization of the results of each generation of the HLOA algorithm, the fitness is calculated before entering the next iteration.

[0073] Next, model identification is performed on the laser output power stabilization system, specifically identifying the internal model and inverse model of the system. The identification results are then compared with the original SVR algorithm and the Particle Swarm Optimization (PSO) SVR algorithm. The comparison results are as follows. Figure 10 and Figure 11 As shown. From Figure 10As can be seen, PSO-SVR and HLOA-SVR show significant improvements in prediction results compared to the unoptimized original SVR algorithm. However, HLOA-SVR performs better than PSO-SVR, reducing the root mean square error of prediction by an order of magnitude. Furthermore, HLOA-SVR exhibits superior "dynamic capture" and "anti-mutation" capabilities. Specifically, when drastic changes occur in the input data around sample points 90, 260, and 450, these points show violent oscillations, indicating that the model failed to keep up with the system changes, resulting in severe lag or prediction failure. However, at the same mutation points, the HLOA-SVR curve remains stable without obvious spikes. This demonstrates that the HLOA-optimized SVR model possesses extremely strong generalization and dynamic tracking capabilities.

[0074] Similarly, corresponding to the inverse model, Figure 11 This demonstrates that the prediction results of the HLOA-SVR algorithm reduce the root mean square error by an order of magnitude compared to the original SVR and PSO-SVR algorithms. Furthermore, at sample points 100, 250, and 450 where data fluctuates, HLOA-SVR also exhibits significant "anti-mutation" capabilities, proving that its unique "multi-strategy optimization" mechanism successfully escapes local traps and finds the global optimal solution for SVR in a huge parameter space. This effectively overcomes the strong nonlinearity and hysteresis characteristics of AOM, laying a solid model foundation for achieving high-precision adaptive internal model control.

[0075] To ensure the generalization ability of the model, five sets of experimental samples were selected and tested using the HLOA-SVR, PSO-SVR, and original SVR methods, respectively. The root-mean-square error (RMSE) and coefficient of determination R of the internal model were measured. 2 The predicted signal-to-noise ratio (SNR) is shown in Table 1 below, and the expression for the predicted SNR is as follows: , Among molecules The total energy representing the effective signal, in the experiment, is the actual change in light intensity output by the AOM, and the denominator is... This represents the total energy of the prediction error.

[0076] Table 1 Analysis of the prediction accuracy of the internal model

[0077] As shown in Table 1, regarding prediction accuracy, while traditional PSO optimization reduced the error by approximately 43%, the HLOA-SVR algorithm proposed in this invention achieved a mean RMSE as low as 0.0044, a further reduction of 87.9% compared to PSO, improving prediction accuracy by an entire order of magnitude. In terms of model purity, HLOA-SVR achieved an average signal-to-noise ratio as high as 44.55 dB, an improvement of approximately 19–23 dB compared to the control group. This indicates that the model has extremely low prediction noise and can accurately reproduce the physical characteristics of the AOM. Regarding stability, even in complex samples where the original model's fit decreased, HLOA maintained a perfect fit of 0.9997, fully validating the effectiveness of its global optimization mechanism.

[0078] Similarly, five groups of samples were selected and tested using three algorithms: HLOA-SVR, PSO-SVR, and the original SVR. The root mean square error (RMSE) and coefficient of determination of the inverse model were obtained. The predicted signal-to-noise ratio (SNR) is shown in Table 2 below.

[0079] Table 2 Analysis of Forward Model Prediction Accuracy

[0080] Table 2 shows a comparative analysis of the experimental data under five different conditions. In terms of accuracy, the HLOA-SVR algorithm exhibits the best performance, with a mean RMSE as low as 0.0058. Compared to the original SVR, the error is reduced by 67.9%; compared to the similar optimization algorithm PSO-SVR, the error is further reduced by 20.3%, demonstrating HLOA's superior ability in parameter optimization. Regarding signal-to-noise ratio (SNR), HLOA-SVR achieves an average SNR of 37.28 dB, an improvement of more than 10 dB compared to the original model. This means that the purity of the model prediction is improved by more than 10 times, effectively eliminating interference and restoring the true physical characteristics of the system. In terms of robustness, as shown in the second group, the fitting accuracy of traditional SVR declines significantly, while HLOA-SVR remains at a high level of around 0.99, demonstrating the algorithm's strong environmental adaptability and stability.

[0081] Regarding HLOA parameter optimization, five sets of data are still selected, with one set used as training samples and the other four sets used as test samples, while setting the population size of the data. Maximum number of iterations Parameter optimization range Under the given conditions, the parameter optimization results obtained through HLOA optimization are shown in Table 3 below.

[0082] Table 3 Parameter optimization results

[0083] The fitness of the system's internal model and inverse model are calculated as follows: Figure 12 As shown, from Figure 12 It can be seen that the HLOA algorithm reaches its minimum fitness in the 27th generation, after which the results tend to plateau. The optimal parameter solution is obtained here.

[0084] Regarding the effectiveness of applying HLOA-SVR-AIMC for stable control of laser output power, such as Figure 13 As shown, this was carried out under actual physical conditions. Experiment on stabilization of continuous laser power. Figure 13 The time-domain evolution of laser output power is shown in both free-running and controlled states. In the free-running phase ( Due to the thermal lensing effect of the AOM crystal and the inherent 1 / f noise of the photodetector, the laser output power exhibits significant random fluctuations, resulting in a relative power fluctuation of approximately ±0.95%. At that time, the proposed HLOA-SVR-AIMC controller seamlessly integrates, that is, as Figure 13 As shown in the enlarged inset, the system quickly and smoothly locks the output power to the target baseline without any non-physical overshoot. Under closed-loop control, severe low-frequency thermal drift is effectively suppressed, compressing long-term steady-state relative fluctuations to only ±0.08%. Compared to free-running, this algorithm improves laser power stability by more than an order of magnitude (nearly 12 times), fully meeting the requirements of high-precision spectral measurements.

[0085] The anti-interference capability of the laser output power stabilization control is evaluated by adding two different step disturbance signals to the system and testing the anti-interference capability of the HLOA-SVR-AIMC-controlled laser output power stabilization system. Specifically: A step noise signal with an amplitude of 0.15 was applied during the time interval; A step noise signal with an amplitude of -0.2 was applied during the time interval, and the beam power output detected by the photodetector after the control signal was adjusted was observed. The input target signal was a step signal with an amplitude of 0.8. The anti-interference test results of the controller are as follows: Figure 14 As shown, the corresponding adaptive parameter change curves are as follows: Figure 15 As shown. From Figure 14 and Figure 15 As can be seen, when interference signals are present, the controller designed in this invention can quickly adjust the modulation voltage of the acousto-optic driver to an appropriate value, rapidly offsetting the beam power shift caused by the interference. Simultaneously, when the noise signal disappears, it will continue to adjust back to the original set desired value. Throughout the entire process, the steady-state error between the system and the target desired value is small, almost maintaining a constant value. Within the order of magnitude, this indicates that the controller has good anti-interference capability.

[0086] Next, the proposed control method is compared with standard PSO-SVR internal model control (PSO-SVR-IMC) and traditional PID control strategies to verify its superiority. Considering the actual operating environment of AOM in a laser power-locked system, three typical disturbance conditions are set in the simulation experiment: EQ \o\ac(○,1) nonlinear operating point step, set value exist (Low transmittance region) and A large step switch is performed between (near saturation region). This setting is intended to simulate the need for wide-range laser power adjustment in methane detection, while verifying the controller's adaptive capability to the nonlinear gain characteristics of the AOM. EQ \o\ac(○,2) transient pulse impact interference, in A pulse interference with an amplitude of 0.2 is introduced at that time, and... Apply a reverse mutation. This setting simulates the mechanical micro-vibrations experienced by the optical platform or the transient electromagnetic shocks experienced by the drive circuitry. The EQ \o\ac(○,3) parameter experiences time-varying drift. An 8% system gain was then introduced for slow drift. This setting corresponds to the thermal lensing effect of the AOM under high-power RF drive and the time-varying characteristics of the crystal refractive index with temperature accumulation. Simulation results are as follows: Figure 16 , Figure 17 And as shown in Table 4.

[0087] Table 4 Comparison of the three controllers

[0088] from Figure 16 , Figure 17 As can be seen from Table 4, in arrive In large-span step jumps, traditional PID methods, due to fixed parameters, are limited. Near-saturation region (physical gain) In the minimum region, it exhibits severe response hysteresis, with a significantly prolonged settling time. Although the PSO-SVR-IMC method has better identification accuracy than traditional PID, it suffers from a fixed filter constant. It is difficult to balance response speed while ensuring no oscillation. In contrast, the HLOA-SVR-AIMC proposed in this invention performs better, as this control method senses local gain in real time. Reduce and automatically adjust the filter constant. This significantly improves the response sensitivity in the saturation region. Experimental data shows that its settling time is reduced by approximately 30% compared to PID control, and it effectively suppresses potential control oscillations in the high-gain region. Under transient pulse interference, the control method proposed in this invention utilizes an internal model architecture to quickly separate external interference, and its recovery time is significantly shorter than that of the control group. After thermally induced gain drift occurs, the PSO-SVR model suffers from insufficient generalization ability of its parameters under complex environments, leading to a gradual accumulation of steady-state residuals with drift. In contrast, the control method proposed in this invention benefits from the constraints of the multi-objective fitness function on physical monotonicity and energy efficiency E. The optimized SVR model exhibits stronger generalization performance. Combined with the internal model feedback correction mechanism, it maintains extremely high locking accuracy even under drift conditions, with RMSE consistently maintained within a certain range. Order of magnitude. In summary, the HLOA-SVR-IMC strategy proposed in this invention not only achieves a deep integration of the mechanism model and data-driven algorithm in theory, but also provides a highly anti-interference and widely adaptable intelligent control scheme at the engineering level, offering new ideas and references for the stabilization control research of other precision optoelectronic instruments.

[0089] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A laser output power stabilization control method based on HLOA-SVR-AIMC, characterized in that, The method includes the following steps: (1) Using the physical mechanism of Bragg diffraction, a nonlinear mapping relationship between diffraction efficiency and driving voltage is established; (2) Based on the regression support vector machine, establish the internal model and inverse model of the laser output power stabilization system, use the inverse model as the controller of the laser output power stabilization system, and introduce an adaptive feedback filter; (3) Set a reference input power signal or voltage signal, and enter the comparator together with the feedback channel signal. Then connect it to the controller, and then enter the module after the controlled object and the internal model are connected in parallel to form an output power signal or voltage signal and a feedback signal. The feedback signal enters the adaptive feedback filter and then enters the comparator to form a closed loop, and obtain the structure diagram of the laser output power stable controlled system. (4) The parameters of the regression support vector machine are globally adaptively searched using the horned lizard optimization algorithm; (5) The fitness function is constructed by integrating the three aspects of prediction accuracy, model complexity and physical logic consistency in a weighted manner to overcome the problems of model overfitting and physical distortion caused by single error optimization in the control process; (6) Select multiple sample sets or datasets from different time periods in the sampling to optimize the parameters in the control process, and use the root mean square error (RMSE) and coefficient of determination (R²) to optimize the parameters. 2 Quantitative testing of predicted signal-to-noise ratio (SNR); (7) Test the adaptability and robustness of the laser output power stabilization system based on the actual reference input power signal or voltage signal and external noise or interference signal, and test the time domain evolution of the laser output power stabilization control system under free operation and control state.

2. The laser output power stabilization control method based on HLOA-SVR-AIMC according to claim 1, characterized in that, The expression for the nonlinear mapping relationship between diffraction efficiency and driving voltage in step (1) is as follows: ； in Indicates the efficiency of diffraction light. Indicates the wavelength of the incident light. This represents the electro-acoustic power conversion gain coefficient of the acousto-optic driver. Indicates the interaction length. Indicates the height of the transducer electrode. Indicates the driving voltage. This represents the acousto-optic quality factor of the crystal.

3. The laser output power stabilization control method based on HLOA-SVR-AIMC according to claim 1, characterized in that, The method for constructing the internal model in step (2) is as follows: collect N sets of input / output datasets of the laser output power stable controlled system. The training sample set D1 is constructed as follows: ， Where y(k) and u(k) are the output and input of the laser output power stabilization system, respectively. For output delay, For input delay, The system output is the total delay between the input and output. As input features, To output labels, and The dimension is , The dimension is Based on the training sample set D1, a regression-type support vector machine is used to train the samples, and the internal model is established as follows: , For internal model functions; The inverse model construction method is as follows: collect N sets of input / output datasets of the laser output power stabilized controlled system. The training sample set D2 is constructed as follows: ， Where y(k) and u(k) are the output and input of the laser output power stabilization system, respectively. For output delay, For input delay, The system output is the total delay between the input and output. Input features for the inverse model, Output labels for the inverse model. The dimension is , The dimension is Based on the training sample set D2, a regression-type support vector machine is used to train the samples, and an inverse model is established as follows: , This is the inverse model function.

4. The laser output power stabilization control method based on HLOA-SVR-AIMC according to claim 1, characterized in that, The adaptive feedback filter in step (2) is: ， Where adaptive filtering parameters , , It is the driving voltage. It is a complex variable. This is a comprehensive coefficient related to the acousto-optic crystal material and optical path structure. Based on the basic filter constant, This is the gain sensitivity adjustment coefficient. This represents the system's maximum local gain.

5. The laser output power stabilization control method based on HLOA-SVR-AIMC according to claim 1, characterized in that, The horned lizard optimization algorithm described in step (4) applies the penalty factor C, kernel width r, and insensitive loss coefficient to the regression support vector machine. Perform a global adaptive search; the regression function of the regression-type support vector machine is: ， In the formula , It is a pair of Lagrange multipliers, and Only a portion of the support vectors are retained at that time. It is the bias constant. It is the kernel width. For support vectors, This represents the number of support vectors.

6. The laser output power stabilization control method based on HLOA-SVR-AIMC according to claim 1, characterized in that, The application of the regression function f(x) approximates the training sample set. , ;when hour, The dataset is Meanwhile, the training sample set satisfies the constraints. ， in These represent the lower and upper slack variables, respectively. Indicates the insensitive loss coefficient; The optimization goal is ， in Indicates the confidence range. This represents the penalty factor.

7. The laser output power stabilization control method based on HLOA-SVR-AIMC according to claim 1, characterized in that, The fitness function in step (5) is: ， Where RMSE is the root mean square error of the training sample set or dataset. The proportion of support vectors to the total number of samples. This is a physical consistency penalty term. If the parameters generated by the horned lizard optimization algorithm cause the model to experience severe oscillations that are not physical fluctuations within the full-range operating voltage of 0.2~0.9V, then this term is assigned a penalty value. E is the total energy consumption, and ω1, ω2, ω3, and ω4 represent the weights of each term.

8. The laser output power stabilization control method based on HLOA-SVR-AIMC according to claim 1, characterized in that, The expression for the total energy consumption E is: ， in It is the power gain coefficient of the driver. The modulation voltage is Δt, which represents the sampling period or sampling time, and the ultrasonic power output by the acousto-optic driver is Δt. With the input modulation voltage The relationship between them is: .

9. The laser output power stabilization control method based on HLOA-SVR-AIMC according to claim 1, characterized in that, In step (6), five sample sets or datasets from different time periods are selected in the sampling to optimize the parameters in the control process, and compared and tested with traditional regression support vector machines and regression support vector machines based on particle swarm optimization.

10. The laser output power stabilization control method based on HLOA-SVR-AIMC according to claim 1, characterized in that, In step (7), computer software is used for simulation or optical platform for experimental testing. The adaptability and robustness of the laser output power stabilization system are tested based on the actual reference input power signal or voltage signal and external noise or interference signal. The HLOA-SVR-AIMC is used to test the time domain evolution of the laser output power stabilization control system under free operation and control states.

Images