High-frequency microwave driving power adaptive control method and system for high-current ECR proton source

Abstract

The application relates to the technical field of microwave control, in particular to a high-frequency microwave driving power adaptive control method and system of a high-current ECR proton source. The method comprises the following steps: collecting electron density, magnetic field intensity, microwave frequency and reflected power in a cavity, and constructing a plasma state data set; constructing a plasma-microwave coupling model, and predicting a critical microwave power threshold value by using a physical information neural network (PINN) model; designing a double closed-loop control mechanism to suppress control chattering; decomposing a microwave signal into a time-frequency domain, combining impedance matching degree and particle swarm optimization to adjust a power envelope curve, and optimizing signal frequency and time distribution; monitoring plasma stability by using a Lyapunov index, and executing a self-healing measure to restore resonance when a loss of lock or an abnormal state is detected; and iteratively updating the PINN model based on PIC and experimental data. The method can improve microwave power control precision, and enhance ECR proton source operation stability and coupling efficiency.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of microwave control, in particular to a high-frequency microwave driving power adaptive control method and system for a high-current ECR proton source. BACKGROUND

[0002] In the high-frequency microwave driving process of a high-current electron cyclotron resonance (ECR) proton source, efficient coupling of microwaves and plasma and stable operation are one of the key technologies. However, the existing technology has insufficient control accuracy of the microwave driving power, and is not sensitive enough to the changes in the plasma state and coupling characteristics, so it is difficult to adapt to complex dynamic conditions. This may not only cause the electron cyclotron resonance state to lose lock, but also cause the coupling efficiency of microwaves and plasma to decrease, thereby affecting the stability and output performance of the proton source. In addition, chattering phenomenon is prone to occur in the control process, further reducing the reliability and efficiency of the system operation.

[0003] In view of this, the application provides a high-frequency microwave driving power adaptive control method and system for a high-current ECR proton source. SUMMARY

[0004] To achieve the above-mentioned purpose, the application provides a high-frequency microwave driving power adaptive control method and system for a high-current ECR proton source, and the specific technical solutions are as follows:

[0005] The high-frequency microwave driving power adaptive control method for a high-current ECR proton source comprises the following steps:

[0006] Real-time acquisition of electron density, magnetic field strength, microwave frequency and reflected power in the ECR plasma cavity is performed by a multi-modal sensor array, and a time-series plasma state data set is constructed after Kalman filtering processing;

[0007] An plasma-microwave coupling model is constructed based on Maxwell's equations, and a physical information neural network (PINN) model is used to predict the critical microwave power threshold under different operating conditions;

[0008] A double-loop control mechanism is designed to adjust the output power of the microwave source, the main loop generates control parameters according to the deviation between the predicted critical microwave power threshold and the actual output power by a fuzzy sliding mode controller, and the secondary loop dynamically optimizes the control parameters by deep reinforcement learning;

[0009] The output power of the microwave source is shaped, the microwave signal is decomposed into the time-frequency domain by fractional Fourier transform, the power envelope curve of the microwave signal is adjusted by a particle swarm optimization algorithm in combination with the plasma impedance matching degree, and the microwave frequency and time distribution structure are optimized;

[0010] The stability of the ECR plasma is monitored by the Lyapunov index, and when the electron cyclotron resonance loses lock or the coupling state is abnormal, a self-healing measure is performed.

[0011] Based on the PIC simulation data and experimental data, the cross-condition features are extracted by using transfer learning, and the weight adjustment amount for updating the physical information neural network PINN model is generated to iteratively optimize the plasma-microwave coupling model.

[0012] Preferably, the electron density, electron temperature, plasma potential information and electric field data in the ECR plasma cavity are obtained; and the magnetic field strength is measured, and the three-dimensional magnetic field distribution of the ECR region is reconstructed;

[0013] The microwave incident power, reflected power and microwave frequency are measured; the coupling state of the microwave and the plasma is evaluated by calculating the reflection coefficient and the standing wave ratio.

[0014] Preferably, a plasma-microwave coupling model is constructed based on the Maxwell equations; and a wave equation in magnetized plasma is derived in combination with the constitutive relation of the plasma;

[0015] A physical information neural network PINN model is used to solve the wave equation in the magnetized plasma, and the network structure of the PINN model is designed as a fully connected deep neural network, including an input layer, a hidden layer and an output layer.

[0016] The input layer receives three-dimensional spatial coordinates, electric field gradients, magnetic mirror field configuration parameters and resonant cavity Q values; the output layer is used to predict the electric field distribution and the critical microwave power threshold; and the predicted critical microwave power threshold is adjusted stably based on the energy balance principle.

[0017] Preferably, a double closed-loop control mechanism is designed to adjust the microwave source output power, and the double closed-loop control mechanism includes using a main loop controller and a secondary loop controller for power control.

[0018] The main loop controller uses a fuzzy sliding mode controller to realize power tracking, and the secondary loop controller dynamically optimizes the control parameters through deep reinforcement learning.

[0019] Preferably, the critical microwave power threshold is obtained after the PINN model calculation, the main loop controller takes the deviation between the predicted critical microwave power threshold and the actual output power as input, and simultaneously defines the power tracking error and the error change rate.

[0020] The power tracking error and the error change rate are used to construct a sliding mode surface function, and a fuzzy logic is introduced to adaptively adjust the switching gain of the main loop controller.

[0021] Preferably, the secondary loop controller uses a deep deterministic policy gradient DDPG algorithm to dynamically optimize the sliding mode surface parameters and the boundary layer thickness; and autonomously learns the optimal combination of control parameters.

[0022] Preferably, the output power of the microwave source is shaped, the microwave signal is decomposed from time domain to time-frequency domain by fractional Fourier transform, and the signal instantaneous frequency and power envelope are jointly analyzed; time-frequency representation of the microwave signal at different rotation angles is obtained, and the order with the highest energy concentration degree is selected as the optimal analysis domain;

[0023] The instantaneous power envelope and the instantaneous frequency of the microwave signal are extracted in the time-frequency domain, and the impedance matching degree is calculated according to the real-time measurement value of the plasma impedance, which is calculated from the reflection coefficient;

[0024] The power envelope optimization objective function is designed based on the impedance matching degree, the particle swarm optimization (PSO) algorithm is used to solve the optimal power envelope curve, and the power envelope is parameterized as a piecewise cubic spline function, and the control points are used as optimization variables.

[0025] Preferably, the frequency structure of the microwave signal is optimized, a controlled frequency modulation is introduced at the center frequency, and the optimal modulation function form is determined by analyzing the plasma dispersion relationship, and the modulation depth is determined by minimizing the reflected power;

[0026] The optimized power envelope and frequency modulation are used to reconstruct the microwave signal by inverse fractional Fourier transform.

[0027] Preferably, the Lyapunov exponent is used to monitor the dynamic stability of the ECR plasma in real time, and the resonance lockout and coupling abnormal state are detected and processed;

[0028] Parameter time series are extracted from the collected time series plasma state data set, and the phase space is reconstructed;

[0029] The electron density data are time sequenced, the time sequenced electron density is used as an observation variable, and the delay coordinate method is used to reconstruct the dimension phase space;

[0030] In the reconstructed phase space, the maximum Lyapunov exponent is calculated to quantify the degree of chaos of the plasma dynamics.

[0031] Preferably, the Lyapunov exponent is used to determine whether the plasma is in a chaotic state, and when the plasma is in a chaotic state, it is determined that the plasma has a resonance lockout;

[0032] The reflection coefficient is used to determine whether a microwave reflection coefficient mutation occurs, and when the microwave reflection coefficient mutation occurs, it is determined that the coupling state of the plasma and the microwave is abnormal.

[0033] Preferably, when the resonance lockout and / or coupling abnormality of the plasma is detected, a multi-stage self-recovery measure is started, and the multi-stage self-recovery measure includes magnetic field fine tuning, frequency scanning, and power soft start;

[0034] The self-recovery measures are executed in priority order, if a single-level measure successfully restores stability, the subsequent measures are terminated, and if all measures fail to restore stability, an alarm is triggered.

[0035] Preferably, based on particle-in-cell (PIC) data and experimental data, a transfer learning method is used to extract common features under different operating conditions, and a physics-informed neural network (PINN) model is continuously optimized.

[0036] The high-frequency microwave driving power adaptive control system of the high-current ECR proton source is used to realize the high-frequency microwave driving power adaptive control method of the high-current ECR proton source, and includes a data acquisition module, a critical threshold prediction module, a double closed-loop control module, a power spectrum shaping module, a self-recovery adjustment module, and an iterative optimization module.

[0037] The data acquisition module acquires the electron density, magnetic field strength, microwave frequency and reflected power in the ECR plasma cavity in real time through a multi-modal sensor array, and constructs a time series plasma state data set after Kalman filtering processing.

[0038] The critical threshold prediction module constructs a plasma-microwave coupling model based on Maxwell's equations, and uses a physics-informed neural network (PINN) model to predict the critical microwave power threshold under different operating conditions.

[0039] The double closed-loop control module designs a double closed-loop control mechanism to adjust the output power of the microwave source, the main loop generates control parameters according to the deviation between the predicted critical microwave power threshold and the actual output power through a fuzzy sliding mode controller, and the secondary loop dynamically optimizes the control parameters using deep reinforcement learning.

[0040] The power spectrum shaping module shapes the output power of the microwave source, decomposes the microwave signal into the time-frequency domain through fractional Fourier transform, adjusts the power envelope curve of the microwave signal by combining the plasma impedance matching degree and using a particle swarm optimization algorithm, and optimizes the microwave frequency and time distribution structure.

[0041] The self-recovery adjustment module monitors the stability of the ECR plasma through the Lyapunov index, and executes self-recovery measures when detecting that the electron cyclotron resonance is lost or the coupling state is abnormal.

[0042] The iterative optimization module extracts cross-condition features using transfer learning based on PIC simulation data and experimental data, generates weight adjustment amounts for updating the physics-informed neural network (PINN) model, and iteratively optimizes the plasma-microwave coupling model.

[0043] The application has the beneficial effects that the application collects key parameters such as electron density and magnetic field strength in real time through a multi-modal sensor, combines Kalman filtering to eliminate noise, accurately constructs a plasma state data set, and provides a reliable data basis for subsequent model prediction and control.

[0044] The application is based on the Maxwell equations and the PINN model, accurately predicts the critical power threshold under different working conditions, helps to realize adaptive adjustment of microwave driving power, optimizes the coupling efficiency of microwave and plasma, and avoids power waste or loss of lock.

[0045] The application constructs a main loop fuzzy sliding mode controller to realize dynamic power regulation, and constructs a secondary loop deep reinforcement learning to optimize control parameters, suppresses control chattering phenomenon, improves control accuracy and response speed, and ensures stable output and rapid adjustment of microwave driving power.

[0046] The application decomposes time-frequency characteristics through fractional Fourier transform, adjusts power envelope curve combined with particle swarm optimization, optimizes frequency and time distribution of microwave signal, enhances plasma impedance matching degree, and improves microwave coupling efficiency and energy utilization rate.

[0047] The application monitors plasma stability in real time through Lyapunov index, quickly detects resonance loss of lock or coupling anomaly, and performs self-healing measures to restore resonance, prevents operation instability, ensures long-term reliable operation of the system, and prolongs the service life of the equipment.

[0048] The application combines PIC simulation data and experimental data, uses transfer learning to extract cross-condition features, continuously optimizes the weights of the PINN model, improves the adaptability and prediction accuracy of the plasma-microwave coupling model, and adapts to variable operating environments. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The high-frequency microwave driving power adaptive control method flowchart of the strong current ECR proton source provided by the application;

[0050] Figure 2 The multi-modal sensor data acquisition and Kalman filtering flowchart provided by the application;

[0051] Figure 3 The PINN model construction and power threshold prediction flowchart provided by the application;

[0052] Figure 4 The double closed loop fuzzy sliding mode control flowchart provided by the application;

[0053] Figure 5 The FRFT power spectrum shaping flowchart provided by the application;

[0054] Figure 6The Lyapunov monitoring and multi-level self-healing flowchart provided by this invention;

[0055] Figure 7 The diagram shows the structure of the high-frequency microwave drive power adaptive control system for the high-current ECR proton source provided by this invention. Detailed Implementation

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0057] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention can also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0058] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0059] Example 1

[0060] Reference Figures 1 to 6 This is the first embodiment of the present invention, such as Figure 1 As shown, an adaptive control method for high-frequency microwave drive power of a high-current ECR proton source is provided.

[0061] Step 1: Electron density, magnetic field strength, microwave frequency, and reflected power within the ECR plasma cavity are acquired in real-time using a multimodal sensor array. After Kalman filtering, a time-series plasma state dataset is constructed. (See attached image) Figure 2 This is a flowchart of the multimodal sensing data acquisition and Kalman filtering process for this step.

[0062] A multimodal sensor array is deployed to comprehensively monitor key physical parameters within the ECR plasma cavity. Several Langmuir probes are arranged along the axial and radial positions of the plasma cavity, forming a three-dimensional measurement grid. The probe spacing can be set to 1 / 4 of the cavity radius to ensure full coverage of the plasma distribution area. It should be noted that the Langmuir probes operate within a scanning voltage range of -50V to +50V, and the scanning period can be set to 10ms. By analyzing the current-voltage characteristic curves collected by the Langmuir probes, the electron density n is extracted. e Electronic temperature T e Plasma potential V p Information and electric field vector E.

[0063] Multiple triaxial Hall sensors are uniformly distributed along the circumference around the cavity. Each sensor can simultaneously measure the three orthogonal components (B, B, and B) of the magnetic field strength B. x B y B z The three-dimensional magnetic field distribution of the ECR region is reconstructed by spatial interpolation of multi-point magnetic field data.

[0064] The incident power P is measured at the microwave feed port using a forward power detector. f The incident power P f This represents the power input from the microwave source to the plasma cavity; the reflected power P is measured by a reverse power detector. r The reflected power P r This indicates the power of microwaves reflected back to the microwave source from the plasma load due to impedance mismatch or other reasons upon reaching the load; the power measurement range is 0-5kW, with a measurement accuracy of ±1%. A real-time spectrum analyzer is also configured to monitor the microwave frequency f. w It is used to capture frequency drift in the range of 2.45 GHz ± 10 MHz.

[0065] The coupling state between microwaves and plasma is evaluated by calculating the reflection coefficient and standing wave ratio.

[0066] The formula for calculating the reflection coefficient is:

[0067]

[0068] The formula for calculating the standing wave ratio is:

[0069] VSWR=(1+|Г|) / (1-|Г|)

[0070] The reflection coefficient ranges from 0 to 1. When Γ=0, it indicates no reflected power, and all input power is absorbed by the plasma; when Γ=1, it indicates that all input power is reflected and no power is absorbed. The standing wave ratio (VSWR) ranges from 1 to ∞. The higher the VSWR, the more severe the reflection and the worse the coupling state.

[0071] The acquired sensor data is processed using the extended Kalman filter algorithm; the state vector x=[n e ,T e ,B,f w ,P f ,P r ,Г,V p ,VSWR] T The state vector x contains 9 key physical quantities; the state transition equation is defined as: x k =f(x k-1 uk-1 )+w k-1 ; where f(·) is a nonlinear state transition function determined from plasma continuity equation and Maxwell equations; u k-1 is a control input vector containing microwave power setpoint and magnetic field coil current; w k-1 is process noise, covariance matrix Q is set according to physical characteristics.

[0072] The observation equation in extended Kalman filter algorithm is represented as: z k = h(x k )+v k ; where h(·) is a nonlinear observation function, v k is measurement noise, covariance matrix R is determined through sensor calibration experiment.

[0073] The Jacobian matrices: F k = , H k = are calculated; the nonlinear state and observation equations are converted to linear form, the prediction and update steps are performed to obtain optimal state estimation, F k represents the partial derivative matrix of state equation with respect to state variable, H k represents the partial derivative matrix of observation equation with respect to state variable, represents the state estimation value at the previous time k-1, represents the state estimation value at the current time k, represents the partial derivative of f(·) with respect to variable x, which is the partial derivative part for calculating F k matrix; represents the partial derivative of h(·) with respect to variable x, which is the partial derivative part for calculating H k matrix.

[0074] The data after Kalman filtering is constructed as a structured time series plasma state dataset , where s is the plasma state vector at the t-th time point in the time series plasma state dataset , N is the number of time point samples, t is the time stamp, p=(r, θ, z) is the spatial coordinate, r represents the radial distance, θ represents the angle, and z represents the height. The state vector s contains the following elements: electron density n e , magnetic field strength B=(B x , B y , B z ), microwave frequency f w , incident power P f , reflected power P r , electron temperature T e , plasma potential V p, the reflection coefficient Г and the standing wave ratio VSWR.

[0075] The acquired plasma data set can be stored in a hierarchical manner. The collected data is stored in the form of a ring buffer, retaining high-frequency sampling data for the last 10 minutes. The historical data collected is compressed in time windows, and complete records for 72 hours are retained for trend analysis. The data update frequency can be set to 100 Hz, and a new state record is generated every 10 ms.

[0076] This step realizes accurate perception and quantitative characterization of the complex dynamic process of ECR plasma by combining multi-modal sensing and intelligent filtering data acquisition methods; the constructed time series plasma state data set contains rich physical information of the plasma, and supports efficient real-time processing and historical analysis through optimized data structure, providing a high-quality data foundation for subsequent power prediction model training and adaptive control strategy optimization.

[0077] Step 2: Construct a plasma-microwave coupling model based on Maxwell's equations, and use a physical information neural network (PINN) model to predict the critical microwave power threshold under different operating conditions; see Figure 3 A flowchart for the PINN model and power threshold prediction for this step is constructed.

[0078] A plasma-microwave coupling model is constructed based on Maxwell's equations to describe the propagation and absorption characteristics of electromagnetic waves in magnetized plasma. Based on Maxwell's equations and the constitutive relations of plasma, the wave equation in magnetized plasma is derived:

[0079]

[0080] where, denotes the gradient, E is the electric field vector, is the microwave angular frequency, c is the speed of light in vacuum, denotes the vector second-order curl of the electric field, is the relative permittivity tensor of the plasma.

[0081] Under ECR conditions, considering the influence of electron cyclotron motion, the relative permittivity tensor is represented as:

[0082]

[0083] where, , is the dielectric constant component perpendicular to the magnetic field direction; , is the off-diagonal component; , is the dielectric constant component parallel to the magnetic field direction; , is the plasma frequency; , is the electron cyclotron frequency, i represents the imaginary part in the imaginary unit, is the electron density, e is the elementary charge, is the vacuum permittivity, is the electron mass, B is the magnetic field strength.

[0084] The wave equation in magnetized plasma is solved by using a physical information neural network (PINN) model. The network structure of the PINN model is designed as a fully connected deep neural network, which includes an input layer, a hidden layer and an output layer. The input layer receives three-dimensional spatial coordinates (x1, y1, z1), an electric field gradient E|, a magnetic mirror field configuration parameter R m , ; wherein B max and B min are the maximum and minimum magnetic field strengths of the magnetic mirror field, respectively, and the Q value of the resonant cavity, the electric field gradient is obtained by central difference calculation from the potential measurements of the Langmuir probe array at multiple adjacent grid points, the Q value of the resonant cavity is obtained by simulation, calibration and online measurement, and is updated in real time during operation; the hidden layer is designed to have 6 layers, each hidden layer contains 128 neurons, and the hyperbolic tangent activation function is used; the output layer is used to predict the electric field distribution E(x1, y1, z1) and the critical microwave power threshold P th .

[0085] The loss function L of the PINN model is designed as the weighted sum of the physical constraint term and the data-driven term:

[0086]

[0087] wherein:

[0088]

[0089] is the partial differential equation residual, o is an index for traversing the partial differential equation discrete points, E o is the value of the electric field, E o is a variable to be solved in the partial differential equation; N PDE is the number of collocation points.

[0090]

[0091] is the boundary condition residual, n is the boundary normal vector, p is an index for traversing the boundary condition points, E p is the value of the electric field under the boundary condition, which is used to calculate the boundary condition loss, is the number of boundary points.

[0092]

[0093] For measuring data residual, q is an index for traversing the measuring data points, N data is the number of experimental measuring points, represents the electric field value predicted by the model, represents the electric field value obtained by experimental measurement; λ1, λ2, λ3 are weight coefficients, and exemplary, the weight coefficients λ1, λ2, λ3 can be set to 1.0, 0.5, 0.8 respectively.

[0094] Through this physically constrained neural network algorithm, complex boundary conditions and nonlinear effects can be flexibly handled under the premise of guaranteeing physical laws, and the calculation efficiency can be significantly improved compared with traditional numerical methods.

[0095] Based on the principle of energy balance, the predicted critical microwave power threshold is adjusted stably, and the power balance equation is defined: ;

[0096] Wherein: ; P ion is the ionization loss power, V plasma is the plasma volume, <σ ion v> is the ionization reaction rate coefficient, E ion is the ionization energy;

[0097] ; P wall is the wall loss power;

[0098] ; is the electron thermal velocity, A wall is the cavity wall area, E wall is the wall loss energy, k B is the Boltzmann constant;

[0099] ; is the radiation loss power, L rad (T e ) is the radiation loss function.

[0100] Exemplarily, when the electron density is 1.0×10 18 m -3 , the electron temperature is 10eV, and the plasma volume is 5.0×10 -4 m 3 , the calculated ionization loss power is about 850W, the wall loss power is about 320W, the radiation loss power is about 45W, and the total critical microwave power threshold is about 1215W.

[0101] The PINN model is trained by the back propagation algorithm and Adam optimizer, with an initial learning rate of 10 -3 , which is gradually reduced to 10 -5 using the cosine annealing strategy. For example, the PINN model uses the Latin hypercube sampling method to generate 10,000 collocation points in the computational domain during the retraining process to ensure uniform spatial distribution; the boundary conditions are applied by uniformly sampling 5,000 points on the surface of the cavity wall. After iterative training, the relative error of the electric field distribution predicted by the PINN model and the finite element method calculation result will be less than 2%, and the calculation time will be shortened from the hour level of the traditional method to the second level.

[0102] This step realizes accurate modeling and fast solving of complex electromagnetic phenomena in ECR plasma by constructing a plasma-microwave coupling model based on physical constraints. The PINN model method embeds physical laws into the neural network architecture, ensuring the physical reasonableness of the prediction results and fully utilizing the powerful fitting ability of deep learning. The predicted critical microwave power threshold provides a key reference value for subsequent adaptive control, ensuring that the microwave source operates near the optimal power point, avoiding plasma extinction caused by insufficient power or mode jump caused by excessive power, thereby significantly improving the operation stability and beam quality of the ECR proton source.

[0103] Step 3: Design a double-loop control mechanism to regulate the output power of the microwave source. The main loop generates control parameters based on the deviation between the predicted critical microwave power threshold and the actual output power using a fuzzy sliding mode controller, and the secondary loop dynamically optimizes the control parameters using deep reinforcement learning to suppress chattering phenomena in the control process; see Figure 3 for the flowchart of this step's double-loop fuzzy sliding mode control.

[0104] Design a double-loop control mechanism to accurately regulate the output power of the microwave source. The main loop uses a fuzzy sliding mode controller to achieve power tracking, and the secondary loop dynamically optimizes the control parameters through deep reinforcement learning. After the PINN model calculation in step 2, the critical microwave power threshold is obtained, and the main loop controller takes the deviation between the predicted critical microwave power threshold P th and the actual output power P out as input.

[0105] Define the power tracking error: ;

[0106] Define the error change rate: ;

[0107] Construct the sliding mode surface function: ;

[0108] wherein c1 and c2 are the sliding surface parameters, and exemplary, the sliding surface parameters c1 and c2 are initially set as c1 = 5, c2 = 1. To suppress the chattering phenomenon of traditional sliding mode control, fuzzy logic is introduced to adaptively adjust the switching gain.

[0109] Two input fuzzy sets are defined, including the power error e p and the sliding surface h m ; the fuzzy set of the power error e p is {negative big (NB), negative medium (NM), negative small (NS), zero (ZO), positive small (PS), positive medium (PM), positive big (PB)}, and the fuzzy set of the sliding surface h m is {negative big (NB), negative medium (NM), zero (ZO), positive medium (PM), positive big (PB)}. The output fuzzy set is the switching gain adjustment coefficient Δk, and the range is {negative big (NB), negative medium (NM), negative small (NS), zero (ZO), positive small (PS), positive medium (PM), positive big (PB)}.

[0110] The triangular membership function is used to describe each fuzzy set, and the input-output mapping relationship is established by 35 fuzzy rules. Exemplarily, when the power error is positive big and the sliding surface is positive big, the switching gain adjustment coefficient output is positive big, indicating that the control action needs to be quickly increased; when the power error is close to zero and the sliding surface is close to zero, the switching gain adjustment coefficient output is zero, maintaining the current control strength. The fuzzy reasoning uses the Mamdani method, and the defuzzification uses the center of gravity method.

[0111] The main loop control law is designed as: ; wherein u main represents the output of the main loop control law, and is the sum of the control quantities.

[0112] ; u eq is the equivalent control term, b is the control gain, is the error acceleration;

[0113] ; u sw is the switching control term, μ is the basic switching gain, Δμ is the fuzzy adjustment amount, sat(˙) is the saturation function, s h is the variable in the fuzzy sliding mode control, is the boundary layer thickness.

[0114] The fuzzy sliding mode control method effectively reduces the high-frequency jitter of the control signal while ensuring robustness, making the microwave source output power more smooth and stable.

[0115] The sub-ring adopts a deep deterministic policy gradient (DDPG) algorithm to dynamically optimize the parameters c1 and c2 of the sliding mode surface and the boundary layer thickness ϕ. An Actor-Critic network architecture is constructed, in which the Actor network is responsible for generating control parameters, and the Critic network evaluates the value of parameter selection.

[0116] The input of the Actor network is the current plasma state vector to be adjusted:

[0117]

[0118] The output is the parameter adjustment action: ;

[0119] where Δc1 and Δc2 are the dynamic adjustment gain parameters of the sliding mode control, is the boundary layer thickness adjustment parameter in the sliding mode control. The Actor network structure is a 4-layer fully connected network, with hidden layer neuron numbers of 400, 300, and 200, respectively, and the activation function uses ReLU. The output layer uses the tanh function to limit the action to the range [-1, 1].

[0120] The input of the Critic network is the state-action pair (s t , a t ), and the output is the Q value function Q(s t , a t ), which represents the long-term return expectation of performing action a t in state s t ; the Critic network structure is similar to the Actor network, but the action vector is spliced with the state features at the second layer;

[0121] The reward function of the Critic network is defined as:

[0122]

[0123] where α1, α2, and α3 are weight coefficients, and exemplary values of the weight coefficients α1, α2, and α3 can be set to 1.0, 0.1, and 10.0, respectively; is the control signal change rate, used to punish chattering; is an indicator function; when the sliding mode surface exceeds the allowed range s max , an additional penalty is given.

[0124] The DDPG algorithm uses an experience replay mechanism, and a replay buffer with a capacity of 10 6 stores historical transition samples (s t , a t , r t , s t+1); each training step performs stochastic sampling of 64 samples from the buffer for mini-batch gradient descent.

[0125] The target network adopts a soft update policy: ; where ξ and ξ' are the main network and target network parameters, respectively, = 0.001 is the soft update coefficient.

[0126] For example, when the initial phase of the sliding mode surface parameters is c1=5, c2=1, and ϕ=0.1, the root mean square value of the power tracking error is 85W, and the amplitude of the control signal chattering is ±120W. After multiple training rounds, the parameters optimized by DDPG are c1=7.3, c2=1.8, and ϕ=0.05, the power tracking error is reduced to 32W, the chattering amplitude is reduced to ±25W, and the control performance is significantly improved.

[0127] This step constructs a double closed-loop control mechanism. The fuzzy sliding mode controller of the main loop ensures fast and accurate tracking of the power set value, and effectively suppresses the inherent chattering problem of traditional sliding mode control through fuzzy logic. The deep reinforcement learning algorithm of the secondary loop learns the optimal combination of control parameters through continuous interaction with the environment, and realizes the continuous optimization of control performance. The two control loops work together to ensure the real-time and robustness of the control, and realize the adaptive adjustment of the parameters, so that the ECR proton source can maintain stable and efficient microwave power output under different operating conditions, significantly improving the consistency of beam quality and the long-term running stability of the ion source.

[0128] Step 4: Shape the output power of the microwave source, decompose the microwave signal into time-frequency domain through fractional Fourier transform, adjust the power envelope curve of the microwave signal through particle swarm optimization algorithm combined with the impedance matching degree of the plasma, optimize the microwave frequency and time distribution structure; refer to Figure 5 , which is the flow chart of FRFT power spectrum shaping for this step.

[0129] The output power of the microwave source is finely shaped to improve the coupling efficiency of microwave energy to the plasma. Fractional Fourier transform (FRFT) is used to decompose the microwave signal from time domain to time-frequency domain, realizing the joint analysis of signal instantaneous frequency and power envelope.

[0130] For the input microwave signal x(t), its α-order fractional Fourier transform X α (u) is defined as:

[0131]

[0132] Where:

[0133]

[0134] Kα (t,u) is the transformation kernel function, where t is the time variable, α=υ'×(π / 2) is the rotation angle, υ' is the fractional order, j is the imaginary unit, and u is the fractional domain coordinate. By scanning within the range of υ'∈[0,2], the time-frequency representation of the microwave signal under different rotation angles is obtained, and the order with the highest energy concentration is selected as the optimal analysis domain.

[0135] Extracting the instantaneous power envelope P of the microwave signal in the time-frequency domain. env (t) and instantaneous frequency f inst (t), the impedance matching degree is calculated based on the real-time measured value of the plasma impedance, which is derived from the reflection coefficient.

[0136] The plasma impedance is expressed as: ;where R p X is the equivalent resistance. p This is the equivalent reactance.

[0137] Define the impedance matching function: ;where R s and X s These are the characteristic resistance and reactance of the microwave transmission line, respectively. When η... m A perfect match is achieved when the value is 1, at which point the microwave energy transmission efficiency is highest.

[0138] Design the power envelope optimization objective function based on impedance matching:

[0139]

[0140] Where T is the optimization time window; the first term in the objective function It is used to maximize power transfer efficiency, the second item Used to limit the rate of power change to avoid plasma disturbance. The smoothing factor is used. The optimal power envelope curve is solved using the Particle Swarm Optimization (PSO) algorithm, which parameterizes the power envelope as a piecewise cubic spline function, with control points used as optimization variables.

[0141] The PSO algorithm initializes several particles, each representing a set of power envelope control points:

[0142]

[0143] Where M is the number of control points, and m represents the particle index;

[0144] The particle velocity update formula is:

[0145]

[0146] in The velocity of the mth particle at the ath iteration, The position of the mth particle at the ath iteration, w is the inertia weight; σ1 and σ2 are acceleration coefficients; r1 and r2 are random numbers in the interval [0, 1]; L best,m is the particle historical optimal position; L best is the global optimal position.

[0147] The particle position is updated as: .

[0148] For example, when the initial power envelope is constant at 1500 W, when the plasma impedance measurement value is Z p = (25 + j15) Ω, the transmission line characteristic impedance is 50 Ω, the initial impedance matching degree is calculated to be 0.64. After 200 iterations of PSO optimization, the time-varying power envelope is obtained, the power is increased to 1800 W in the resonance region and decreased to 1200 W in the non-resonance region, the average impedance matching degree is increased to 0.85, and the microwave coupling efficiency is increased by about 33%.

[0149] At the same time, the frequency structure of the microwave signal is optimized, and a controlled frequency modulation is introduced near the center frequency f0= 2.45 GHz: ; Where Δf is the frequency modulation depth, h(t) is the modulation function. By analyzing the plasma dispersion relation, the optimal modulation function is determined to be a sawtooth wave form, and the period is synchronized with the plasma density fluctuation period. The modulation depth Δf is determined by minimizing the reflected power, and the typical value is in the range of ± 5 MHz.

[0150] The optimized power envelope P * env (t) and the frequency modulation f * (t) are reconstructed into a microwave signal through inverse fractional Fourier transform:

[0151]

[0152] Where IFrFT α represents the inverse fractional Fourier transform, τ is the integral variable, and represents the time from the starting time 0 to the time t. The reconstructed signal maintains the phase continuity of the original signal, avoiding the spectral broadening caused by sudden changes.

[0153] The step realizes the joint optimization of the time-frequency characteristics of the microwave signal by a power spectrum shaping method based on fractional Fourier transform. The optimized power envelope is adaptively adjusted according to the dynamic change of the plasma impedance, ensuring that a high energy coupling efficiency is maintained at each moment. The introduction of frequency modulation broadens the interaction bandwidth between the microwave and the plasma, and enhances the adaptability to the fluctuations of the plasma parameters. The power spectrum shaping process not only improves the utilization rate of microwave energy, but also avoids the impact on the stability of the plasma by smooth power and frequency changes, providing an important guarantee for the efficient and stable operation of the ECR proton source.

[0154] Step 5: Monitor the stability of the ECR plasma by Lyapunov exponent, and when the electron cyclotron resonance is detected to be out of lock or the coupling state is abnormal, execute self-healing measures to restore the resonance condition and suppress the operation instability; see Figure 6 The Lyapunov monitoring and multi-stage self-healing flowchart for this step.

[0155] The dynamics stability of the ECR plasma is monitored in real time by Lyapunov exponent, and the resonance out of lock and the abnormal coupling state are detected and processed in time. The time series of key parameters are extracted from the time series plasma state data set collected in step 1, and the phase space reconstruction is constructed.

[0156] The electron density n e is time-series processed, and the time-series processed electron density n e (t) is taken as the main observation variable, and the delay coordinate method is used to reconstruct the dimensional phase space:

[0157]

[0158] Where υ is the time delay, which is determined by the first minimum value of the mutual information function; is the embedding dimension, which is determined by the false nearest neighbor method. In the reconstructed phase space, the maximum Lyapunov exponent λ max is calculated to quantify the degree of chaos of the plasma dynamics.

[0159] The Wolf algorithm is used to track the divergence rate of adjacent orbits in the phase space:

[0160]

[0161] Where d u is the distance between the two orbits before the u-th evolution, d u ’ is the distance after evolution, and U is the evolution step number. When λ max > 0, it indicates that the plasma is in a chaotic state, and stability measures need to be taken immediately.

[0162] Establish stability criterion, define critical Lyapunov exponent threshold λ cr When λ max > λ cr , determine that the electron cyclotron resonance is lost. At the same time, monitor the sudden change of microwave reflection coefficient Г, when |ΔГ / Δt| > Г th , where Г th is the threshold of reflection coefficient change rate, determine that the coupling state is abnormal.

[0163] When resonance loss or coupling abnormality is detected, immediately start three-level self-healing measures. The first level is magnetic field fine tuning, which re-matches the electron cyclotron resonance condition by accurately adjusting the magnetic field coil current. Calculate the corresponding resonance magnetic field strength B w : ; where m e is the electron mass, and e is the basic charge.

[0164] Get the current actual magnetic field strength B act , calculate the deviation: .

[0165] Adjust the magnetic field coil current through a proportional-integral controller:

[0166]

[0167] where I0 is the initial coil current, K p and K o are the proportional and integral gains respectively. The magnetic field adjustment rate is limited within dB / dt < 0.5 mT / s to avoid plasma disturbance caused by too fast adjustment.

[0168] For example, when the microwave frequency is 2.45 GHz, if the calculated resonance magnetic field strength is 87.5 mT and the actual measured magnetic field is 85.2 mT, the deviation is 2.3 mT. Set K p = 0.8 A / mT, K o = 0.2 A / (mT·s), and the initial current I0 = 100 A, after 3 seconds of adjustment, the coil current increases to 101.84 A, and the magnetic field strength recovers to within the range of 87.5 ± 0.1 mT.

[0169] The second level is frequency scanning, when the magnetic field fine tuning fails to restore resonance, dynamically adjust the microwave frequency to find a new coupling point, perform linear frequency scanning around the center frequency f0:

[0170]

[0171] where v f is the frequency scanning rate; t startThe scan start time is t0. The scan range is limited to f0±10MHz. The reflected power P is monitored in real time r When P r When a significant drop occurs (e.g. more than 30%), the current frequency is locked as the new operating frequency.

[0172] The frequency adjustment is realized by a voltage-controlled oscillator (VCO), and the control voltage and frequency offset relationship is: ; where V0 is the center voltage, k VCO is the tuning sensitivity.

[0173] The third level is power soft start. When the current two levels are ineffective, the power reduction and gradual recovery program is executed;

[0174] First, the microwave power is quickly reduced to a safe level , where is the rated power, is the microwave power adjustment coefficient, which keeps the plasma from extinguishing.

[0175] Then the power is gradually recovered according to the exponential law:

[0176]

[0177] , where is the target power, is the recovery time constant, which can be set to 5 seconds. The Lyapunov index is continuously monitored during the power recovery process. If unstable signs appear again, the power rise is immediately stopped and the current level is maintained.

[0178] The three-level self-healing measures are executed in priority order. If a single-level measure successfully restores stability, subsequent measures are terminated. If all measures fail to restore stability, an alarm is triggered and the system switches to safe mode operation. The status information of the entire self-healing process is recorded in the log file, including the triggering time, Lyapunov index value, measures taken, and recovery effect, providing data support for subsequent fault analysis and control strategy optimization.

[0179] This step realizes active maintenance of ECR plasma stability through real-time monitoring of Lyapunov index and the synergistic effect of multi-level self-healing mechanism. This method not only quickly detects unstable states such as resonance lockout and coupling abnormalities, but also automatically selects appropriate recovery measures according to specific conditions, significantly reducing the probability of plasma extinction and mode jump. Compared with traditional passive protection mechanisms, the active stability control method extends the average fault-free operation time of the plasma by several times, significantly improving the operation reliability of the ECR proton source and the continuity of the beam output.

[0180] Step 6: Based on PIC simulation data and experimental data, transfer learning is used to extract cross-condition features, and the weight adjustment amount for updating the physical information neural network (PINN) model is generated to iteratively optimize the plasma-microwave coupling model.

[0181] Based on PIC simulation data and experimental data, transfer learning is used to extract common features under different operating conditions, and the physical information neural network (PINN) model is continuously optimized. A comprehensive data set containing multiple conditions is constructed, and PIC data simulation covers the plasma evolution process under different gas pressures (Pa) to microwave power (500W to 5000W) and magnetic field configuration. The particle simulation uses a full electromagnetic PIC code, with a particle number of superparticles and a time step that satisfies the Courant condition , where is the spatial grid size and c is the speed of light. Multiple microwave periods are simulated for each condition, and the electromagnetic field distribution, particle phase space distribution, and macroscopic plasma parameters are output.

[0182] The experimental data comes from the running records of the ECR proton source, specifically the historical data accumulated during operation. The experimental data and the corresponding PIC simulation data are time-aligned and spatially interpolated to construct a paired data set: ; where is the simulation input feature, is the experimental input feature, is the target output (critical microwave power), is the total number of samples.

[0183] A domain adaptive neural network architecture is designed, including a shared feature extractor , simulation domain feature extractor , experimental domain feature extractor , and predictor . The shared feature extractor uses a residual network structure to extract cross-domain invariant features. The domain-specific extractors learn the unique patterns of simulation and experimental data, respectively. Feature fusion is achieved through an attention mechanism: ;

[0184] where is the fused feature, and are attention weights, which are dynamically calculated through a soft attention mechanism, according to the source of the input data or . For simulation input features , select ; for experimental input features Then Select .

[0185] For example, in the initial training stage, 80% of the PIC simulation data is used to pre-train the network to obtain a basic model. Then 20% of the experimental data is introduced for fine-tuning. If the accuracy of the domain discriminator decreases from the initial 95% to 55%, it indicates that the domain-invariant features have been successfully learned.

[0186] Based on the features extracted by transfer learning, the weight adjustment amount of the PINN model is generated; and a weight update strategy is defined: ; wherein is the updated weight of the PINN model, is the current weight of the PINN model in step 2, is the adjustment amount calculated by transfer learning, is the learning rate, and an adaptive adjustment strategy is adopted. The weight adjustment amount is determined by minimizing the degree of violation of physical constraints to ensure that the updated PINN model still satisfies the Maxwell equations.

[0187] A triggering mechanism for updating the PINN model is established. When the cumulative runtime limit is reached or the operating condition changes significantly, the transfer learning process is automatically started. After the transfer learning is updated, A / B testing is performed to verify the performance of the new PINN model in the parallel running test channel.

[0188] Through this model updating mechanism based on transfer learning, deep integration of theoretical simulation and actual running data is achieved. The microscopic physical insights provided by PIC simulation and the macroscopic behavior reflected by experimental data complement each other, enabling the PINN model to accurately capture the complex dynamic characteristics of ECR plasma.

[0189] Embodiment 2

[0190] Referring to Figure 7 , the second embodiment of the present application provides a high-frequency microwave driving power adaptive control system for a high-current ECR proton source.

[0191] The system includes a data acquisition module, a critical threshold prediction module, a double closed-loop control module, a power spectrum shaping module, a self-healing adjustment module, and an iterative optimization module.

[0192] The data acquisition module acquires the electron density, magnetic field strength, microwave frequency and reflected power in the ECR plasma cavity in real time through a multi-modal sensor array, and constructs a time series plasma state data set after Kalman filtering processing.

[0193] The critical threshold prediction module is configured to construct a plasma-microwave coupling model based on Maxwell equations, and to predict the critical microwave power threshold under different operating conditions by using a physical information neural network (PINN) model.

[0194] The double closed-loop control module is configured to adjust the output power of the microwave source by using a double closed-loop control mechanism, a main loop is configured to generate a control parameter according to a deviation between the predicted critical microwave power threshold and the actual output power by using a fuzzy sliding mode controller, and a secondary loop is configured to dynamically optimize the control parameter by using deep reinforcement learning to suppress chattering in the control process.

[0195] The power spectrum shaping module is configured to shape the output power of the microwave source, to decompose the microwave signal into a time-frequency domain by using fractional Fourier transform, to adjust a power envelope curve of the microwave signal by using a particle swarm optimization algorithm in combination with plasma impedance matching degree, and to optimize a microwave frequency and time distribution structure.

[0196] The self-healing adjustment module is configured to monitor the stability of the ECR plasma by using a Lyapunov index, to execute a self-healing measure to restore resonance conditions and suppress operating instability when detecting that the electron cyclotron resonance is lost or the coupling state is abnormal.

[0197] The iterative optimization module is configured to extract cross-condition features by using transfer learning based on PIC simulation data and experimental data, to generate a weight adjustment amount for updating the physical information neural network (PINN) model, and to iteratively optimize the plasma-microwave coupling model.

[0198] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other manners. The described device embodiments are merely illustrative. For example, the division of the units is only a logical function division. Other division manners can be used in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0199] The embodiments of the present application are described above with reference to the drawings, but the present application is not limited to the above specific embodiments. The above specific embodiments are merely illustrative, rather than restrictive, and a person of ordinary skill in the art can make changes, modifications, replacements and variations to the above embodiments without departing from the purpose of the present application and the scope protected by the claims, which are all within the protection of the present application.

Claims

1. A method for adaptive control of high-frequency microwave drive power of a high-current ECR proton source, characterized in that, include: The electron density, magnetic field strength, microwave frequency and reflected power in the ECR plasma cavity are collected in real time by a multimodal sensor array, and a time-series plasma state dataset is constructed after Kalman filtering. A plasma-microwave coupling model was constructed based on Maxwell's equations, and the critical microwave power threshold under different operating conditions was predicted using the PINN physical information neural network model. A dual closed-loop control mechanism is designed to adjust the output power of the microwave source. The main loop generates control parameters based on the deviation between the predicted critical microwave power threshold and the actual output power through a fuzzy sliding mode controller. The secondary loop uses deep reinforcement learning to dynamically optimize the control parameters. The output power of the microwave source is shaped, and the microwave signal is decomposed into the time and frequency domain by fractional Fourier transform. Combined with the plasma impedance matching degree, the power envelope curve of the microwave signal is adjusted by particle swarm optimization algorithm to optimize the microwave frequency and time distribution structure. The stability of ECR ​​plasma is monitored by Lyapunov index, and self-healing measures are implemented when electron cyclotron resonance lock-up or abnormal coupling state is detected. Based on PIC simulation and experimental data, transfer learning is used to extract cross-condition features, generate weight adjustment values ​​for updating the PINN neural network model of physical information, and iteratively optimize the plasma-microwave coupling model.

2. The high-frequency microwave drive power adaptive control method for a high-current ECR proton source according to claim 1, characterized in that, Acquire electron density, electron temperature, plasma potential information, and electric field data within the ECR plasma cavity; measure magnetic field strength and reconstruct the three-dimensional magnetic field distribution in the ECR region; The microwave incident power, reflected power, and microwave frequency are measured; the coupling state between microwave and plasma is evaluated by calculating the reflection coefficient and standing wave ratio.

3. The high-frequency microwave drive power adaptive control method for a high-current ECR proton source according to claim 2, characterized in that, A plasma-microwave coupling model was constructed based on Maxwell's equations; and the wave equation in magnetized plasma was derived by combining the constitutive relation of plasma. The wave equation in magnetized plasma is solved using the Physical Information Neural Network (PINN) model. The network structure of the PINN model is designed as a fully connected deep neural network, which includes an input layer, a hidden layer, and an output layer. The input layer receives three-dimensional spatial coordinates, electric field gradient, magnetic mirror field configuration parameters, and resonant cavity Q value; the output layer is used to predict the electric field distribution and critical microwave power threshold; and the predicted critical microwave power threshold is steadily adjusted based on the energy balance principle.

4. The high-frequency microwave drive power adaptive control method for a high-current ECR proton source according to claim 3, characterized in that, Design a dual closed-loop control mechanism to adjust the output power of the microwave source. The dual closed-loop control mechanism includes power control using a main loop controller and a secondary loop controller. The main loop controller adopts a fuzzy sliding mode controller, and the sub-loop controller dynamically optimizes the control parameters through deep reinforcement learning.

5. The high-frequency microwave drive power adaptive control method for a high-current ECR proton source according to claim 4, characterized in that, After calculation using the PINN model, the critical microwave power threshold is obtained. The main loop controller uses the deviation between the predicted critical microwave power threshold and the actual output power as input, and defines the power tracking error and the error change rate. A sliding mode surface function is constructed using power tracking error and error change rate, and fuzzy logic is introduced to adaptively adjust the switching gain of the main loop controller.

6. The high-frequency microwave drive power adaptive control method for a high-current ECR proton source according to claim 5, characterized in that, The sub-loop controller uses the Deep Deterministic Strategy Gradient DDPG algorithm to dynamically optimize the sliding surface parameters and boundary layer thickness; it autonomously learns the optimal combination of control parameters.

7. The high-frequency microwave drive power adaptive control method for a high-current ECR proton source according to claim 6, characterized in that, The output power of the microwave source is shaped, and the microwave signal is decomposed from the time domain to the time-frequency domain using fractional Fourier transform. The instantaneous frequency and power envelope of the signal are jointly analyzed. The time-frequency representation of the microwave signal under different rotation angles is obtained, and the order with the highest energy concentration is selected as the optimal analysis domain. The instantaneous power envelope and instantaneous frequency of the microwave signal are extracted in the time-frequency domain, and the impedance matching degree is calculated based on the real-time measurement value of the plasma impedance, which is derived from the reflection coefficient. The objective function for power envelope optimization is designed based on impedance matching degree. The optimal power envelope curve is solved by particle swarm optimization (PSO) algorithm. The power envelope is parameterized as a piecewise cubic spline function, and the control points are used as optimization variables.

8. The high-frequency microwave drive power adaptive control method for a high-current ECR proton source according to claim 7, characterized in that, The frequency structure of the microwave signal is optimized by introducing controlled frequency modulation at the center frequency, and the optimal modulation function is determined by analyzing the plasma dispersion relationship. The modulation depth is determined by minimizing the reflected power. The optimized power envelope and frequency modulation are used to reconstruct the microwave signal through inverse fractional Fourier transform.

9. The high-frequency microwave drive power adaptive control method for a high-current ECR proton source according to claim 8, characterized in that, Real-time monitoring of the dynamic stability of ECR ​​plasma using the Lyapunov index to detect and handle resonance lockout and coupling anomalies; The time series of parameters are extracted from the collected time-series plasma state dataset, and the phase space is reconstructed. The electron density data is time-series analyzed, and the time-series electron density is used as the observation variable. The delayed coordinate method is then used to reconstruct the 3D phase space. In the reconstructed phase space, the maximum Lyapunov exponent is calculated to quantify the degree of chaos in plasma dynamics.

10. The high-frequency microwave drive power adaptive control method for a high-current ECR proton source according to claim 9, characterized in that, The Lyapunov index is used to determine whether the plasma is in a chaotic state. When the plasma is in a chaotic state, it is determined that the plasma has experienced electron cyclotron resonance lock-up. The microwave reflection coefficient is used to determine whether there is a sudden change in the microwave reflection coefficient. When a sudden change occurs in the microwave reflection coefficient, it is determined that the coupling state between the plasma and the microwave is abnormal.

11. The high-frequency microwave drive power adaptive control method for a high-current ECR proton source according to claim 10, characterized in that, When a plasma resonance lockout and / or coupling anomaly is detected, a multi-level self-healing mechanism is initiated, which includes magnetic field fine-tuning, frequency scanning, and power soft-start. Self-healing measures are executed in order of priority. If a single-level measure successfully restores stability, subsequent measures are terminated; if all measures fail to restore stability, an alarm is triggered.

12. The high-frequency microwave drive power adaptive control method for a high-current ECR proton source according to claim 11, characterized in that, Based on particle simulation (PIC) data and experimental data, a transfer learning method was used to extract common features under different operating conditions, and the physical information neural network (PINN) model was continuously optimized.

13. A high-frequency microwave drive power adaptive control system for a high-current ECR proton source, used to implement the high-frequency microwave drive power adaptive control method for the high-current ECR proton source according to any one of claims 1 to 12, characterized in that, include: The system includes a data acquisition module, a critical threshold prediction module, a dual closed-loop control module, a power spectrum shaping module, a self-healing adjustment module, and an iterative optimization module. The data acquisition module collects the electron density, magnetic field strength, microwave frequency and reflected power in the ECR plasma cavity in real time through a multimodal sensor array, and constructs a time-series plasma state dataset after Kalman filtering. The critical threshold prediction module constructs a plasma-microwave coupling model based on Maxwell's equations and uses the PINN physical information neural network model to predict the critical microwave power threshold under different operating conditions. The dual-loop control module is designed with a dual-loop control mechanism to adjust the output power of the microwave source. The main loop generates control parameters based on the deviation between the predicted critical microwave power threshold and the actual output power through a fuzzy sliding mode controller. The secondary loop uses deep reinforcement learning to dynamically optimize the control parameters. The power spectrum shaping module shapes the output power of the microwave source, decomposes the microwave signal into the time and frequency domain through fractional Fourier transform, and adjusts the power envelope curve of the microwave signal through particle swarm optimization algorithm in combination with plasma impedance matching degree to optimize the microwave frequency and time distribution structure. The self-healing adjustment module monitors the stability of the ECR plasma through the Lyapunov index. When an electron cyclotron resonance lock-up or abnormal coupling state is detected, it executes self-healing measures. The iterative optimization module, based on PIC simulation data and experimental data, uses transfer learning to extract cross-condition features, generates weight adjustment values ​​for updating the physical information neural network PINN model, and iteratively optimizes the plasma-microwave coupling model.

Images