CN113221432B - A dynamic prediction method for ion thruster grid life based on artificial intelligence - Google Patents

Abstract

The present invention discloses a method for dynamically predicting the life of an ion thruster gate based on artificial intelligence. First, a three-dimensional numerical simulation model of the gate is used to simulate and analyze multiple typical operating conditions of the ion thruster gate system, and then a database under different parameter conditions is established based on the simulation results. Finally, the database is used to train a neural network to form a mapping between multiple gate parameters and gate wall corrosion rate. This mapping can be used to predict the change in gate life in real time according to the change in gate operating parameters. This method is more in line with the actual situation than the existing method of estimating the gate life based on a single operating condition, and is therefore more reliable and reasonable, and solves the problem of excessive calculation of multi-operating condition and multi-parameter simulation. Ultimately, it can meet the life estimation needs of satellites and other spacecraft for ion thrusters under a wide range of variable thrust.

Description

A dynamic prediction method for ion thruster grid life based on artificial intelligence

Technical Field

The invention belongs to the technical field of ion thrusters and relates to a method for dynamically predicting the life of an ion thruster grid based on artificial intelligence.

Background Art

In order to meet the satellite's demand for a wide range of variable thrust, the ion thruster grid system needs to work in a multi-mode real-time switching state during actual operation. At this time, the grid system parameters, such as upstream plasma density, aperture, voltage and other parameters, also change in real time. Therefore, the traditional method of estimating the grid life using fixed operating condition simulation will produce a large error compared to the actual operating conditions. On the other hand, the real-time dynamic changes in parameters caused by the multi-mode switching of the grid pose a huge challenge to the numerical simulation of the grid system.

Depending on whether the acceleration gate has direct current interception, the operation state of the gate can be roughly divided into over-focusing, normal focusing and under-focusing. The simulation grids required for each operation state are very different, and usually increase exponentially with the upstream plasma density. Therefore, if a single example is used to directly simulate the actual multi-mode operation process of the gate, it cannot be achieved with the existing computer capabilities. For such problems, the commonly used idea is iterative correction, that is, first calculate the gate operation to a steady state based on a set of initial parameters, then correct the gate parameters according to the steady-state results of the initial parameters, and then continue the simulation calculation under the corrected parameters until the total iteration time is equal to the actual operation time of the gate. However, this idea still meets the problem of excessive computational complexity. Although a single example meets the computer computing power requirements under this method, the amount of computation required to complete the entire iteration is too large.

Therefore, based on the iterative calculation of some representative working conditions, we can consider directly calculating the representative working conditions of the gate, and then fitting the functional relationship between different gate parameters and gate life based on the representative working condition results. However, the relationship between gate parameters and gate life is highly nonlinear, and there are many gate parameters. It is difficult to find suitable fitting coefficients using commonly used linear fitting or polynomial fitting. At the same time, the emerging artificial intelligence method is a means that is particularly suitable for finding multi-parameter, nonlinear functions.

Summary of the invention

In order to overcome the deficiencies in the prior art, the present invention solves the problem of dynamic prediction of the gate life of an ion thruster. The present invention proposes to establish a dynamic prediction model of gate life by combining a gate simulation model and an artificial intelligence method. The specific idea is: first, using the existing gate simulation model, simulate representative working conditions to obtain simulation results of representative working conditions. Then, a database is formed using the simulation results of all representative working conditions, and a neural network is trained using the database to obtain a functional relationship between gate parameters and gate corrosion rate (i.e., gate life). Finally, the trained neural network is used to predict the gate life under any combination of working conditions.

The technical solution of the present invention is: a method for dynamically predicting the life of an ion thruster grid based on artificial intelligence,

First, a gate simulation program based on the macro particle-Monte Carlo collision (PIC-MCC) algorithm is established. Secondly, a post-processing program for solving the gate corrosion rate is established to solve the corrosion rate distribution and average corrosion rate of the gate end face and the hole wall. Thirdly, a database is formed by each typical operating condition parameter, the failure limit parameter and its corresponding gate wall corrosion rate. Then, a neural network model based on the BP neural network algorithm is established, and the neural network is trained using the established database. The trained neural network can establish a multi-parameter mapping for different operating conditions. Then, for a single-mode working gate, the input parameters of the PIC-MCC model can be repeatedly called to continuously correct the input parameters of the PIC-MCC model to obtain the input parameters at the final moment, and then the gate life performance parameters at the end of the life are obtained by simulation calculation and the gate life is given. Finally, in the multi-mode wear etching prediction, different mappings are used in different modes to iterate to achieve mode switching, and the final life of the gate is obtained by adding gate failure criteria during the entire iterative process.

It is characterized by comprising the following steps:

Step 1, establishing a gate simulation model based on a macro particle-Monte Carlo collision algorithm;

Step 2, establishing a post-processing model for solving the gate corrosion rate, solving the corrosion rate distribution and average corrosion rate of the gate end surface and the hole wall;

Step 3, forming a database from typical operating parameters, failure limit parameters and their corresponding gate wall corrosion rates;

Step 4, establishing a neural network model based on the BP neural network algorithm, using the database established in step 3 to train the neural network, and establishing multi-parameter mapping for different working conditions with the trained neural network;

Step 5: For the single-mode working gate, the input parameters based on the macro-particle-Monte Carlo collision model are continuously corrected by repeatedly calling the mapping to obtain the input parameters at the final moment, and then the gate life performance parameters at the end of the life are obtained by simulation calculation and the gate life is given;

Step 6, in the multi-mode wear etching prediction, different mappings are used to iterate through different modes to achieve mode switching, and the final gate life is obtained by adding gate failure criteria in the entire iterative process.

The beneficial effects of the present invention compared with the prior art are:

In this algorithm, on the basis of iterative calculation of some representative working conditions, the currently emerging artificial intelligence method is used to obtain the nonlinear function between the multi-parameters of the gate and the corrosion rate of the gate wall. The problem of huge computational complexity of multi-mode, multi-condition and multi-parameter simulation of the gate is solved, while the accuracy of multi-mode gate life estimation is greatly improved, and the gate life at any time, under any working condition or any combination of working conditions can be estimated in real time. After evaluation and verification, this simulation method is considered to be effective in radial dynamic prediction of the life of the ion thruster gate, and after comparison with the experimental results, the calculation accuracy is good, which has important guiding significance for the performance verification of the thruster gate in the early stage of design, as well as the optimization of thruster parameters in the later stage, the direction of experimental parameter adjustment and the design range. This design method is suitable for the early design of the ion thruster gate system and the later experimental guidance.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a flow chart of the PIC-MCC algorithm.

Figure 2 is the BP neural network algorithm flow.

FIG. 3 is a schematic diagram of finding the mapping f using a neural network.

FIG. 4 is a schematic diagram of iteration using the mapping f.

FIG. 5 is a schematic diagram of a mode switching simulation scheme.

FIG. 6 is a schematic diagram of adding failure criteria during the iteration process.

DETAILED DESCRIPTION

The present invention is a dynamic prediction method for the life of an ion thruster grid based on artificial intelligence. The main goal of the present invention is to simulate and analyze the wear life of a multi-mode ion thruster grid, and form a set of simulation tools that can dynamically predict the life of a multi-mode ion thruster grid. In the actual working process of the grid, the upstream plasma density, neutral atom density, grid voltage, aperture, hole spacing, grid thickness and grid spacing are all in dynamic change. And in each working mode, the grid may be in completely different working conditions. From the perspective of numerical simulation, its essence is a feedback correction process, that is, a feedback correction is used to check the grid input parameters to accurately and real-time reflect the grid operation state. Therefore, in view of the above problem, the idea of the present invention is to iteratively correct the grid wear prediction model based on the dynamic feedback of the neural network algorithm. In order to obtain the grid corrosion rate that can reflect any parameter, the neural network needs to be trained. Therefore, firstly, a few samples are obtained by grid program simulation, and a database is established; then the database is imported into the neural network system to train the neural network; finally, the trained neural network is the mapping. For a single mode, a few samples can be generated by numerical simulation, and then the mapping can be found by training the neural network; then the mapping is repeatedly called, the input parameters are continuously corrected, the input parameters at the final moment are obtained, and then the simulation calculation is performed. For multi-mode cutting, we first obtain the mapping of different modes, and then iterate using different mappings for different modes. For the dynamic model of life prediction, we can solve it by adding gate failure criteria in the entire iterative process.

The implementation steps are as follows:

(1) First, as shown in Figure 1, a gate simulation model based on the macroparticle-Monte Carlo collision (PIC-MCC) algorithm is established. This model consists of two parts: a model that simulates the movement of ion beams in the optical system and a model that simulates the generation of charge exchange (CEX) ions. The first part of the program is used to simulate the extraction process of beam ions. During this process, there are very few collisions between ions, which can be ignored. The simulation ends when the total number of ions reaches a stable state. During the simulation, the ratio of divalent ions to monovalent ions is calculated according to the following formula:

In the formula, are the divalent ion and monovalent ion densities, respectively; I _d , I _b are the discharge current and beam current, respectively; v ₀ , _vi are the atomic and ion velocities, respectively; T _a , η _c are the gate transmittance and Clauin coefficient of the atom, respectively; σ ₊₊ , σ ₊ are the divalent ion ionization collision cross section and the monovalent ion ionization collision cross section, respectively; γ is the ionization rate. In addition, neutral atoms are uncharged, so the distribution of neutral atoms will not affect the electric field in the simulation area. This embodiment assumes that neutral atoms are uniformly distributed in the calculation area. The second part of the model is used to simulate the CEX ion distribution. Since the number of CEX ions is a very small amount relative to the beam ions, it can be considered that the CEX ion density has no effect on the overall potential distribution. The beam ion distribution and the neutral atom distribution are used to simulate the generation and movement of charge exchange ions in the simulation area. In the first few hundred hours of the gate working, it can be considered that the amount of corrosion of the gate wall and the hole wall is a small amount, and the corrosion process has no effect on the overall potential distribution. The corrosion depth calculation model mainly uses the relevant physical quantities of beam ions and CEX ions to calculate the corrosion size of the accelerating grid hole wall and downstream surface.

(2) Establish a post-processing model to solve the gate corrosion rate, and solve the corrosion rate distribution and average corrosion rate of the gate end surface and hole wall. The definition of corrosion rate is obtained to obtain its expression:

Where _RE is the corrosion rate, J is the ion current density on the gate wall, Y is the ion sputtering yield, q is the ion charge, _Mg is the atomic mass of the gate material, and _ρg is the density of the gate material. According to the corrosion rate expression, it can be seen that the _RE corrosion rate of the gate wall can be obtained by statistically obtaining the ion current density hitting the gate surface using the PIC-MCC gate simulation model.

(3) Sorting out typical working conditions and corresponding working parameters of the multi-mode of the gate, mainly including: geometric parameters, such as the thickness of the accelerating gate, the aperture of the accelerating gate, the thickness of the decelerating gate, the aperture of the decelerating gate, the screen-accelerating distance, the acceleration-decelerating distance; electrical parameters, such as the screen gate voltage, the accelerating gate voltage, the decelerating gate voltage; plasma parameters, such as the upstream plasma density, the neutral atom density, and the anode voltage (which can be characterized as the proportion of divalent ions). At the same time, the gate failure limit parameters are estimated. For example, the failure limit radius of the accelerating gate electron backflow can be calculated according to the following formula:

Where V _N is the net accelerating voltage, which is the sum of the screen grid voltage and the relative screen grid potential of the plasma in the ionization chamber; _VT is the total accelerating voltage, which is the difference between the net accelerating voltage and the accelerating grid voltage; l _g , _rs , _ta and _ra are the grid spacing, screen grid hole radius, accelerating grid thickness and accelerating grid hole radius, respectively.

Then, the PIC-MCC gate simulation model is used to simulate each typical operating condition parameter and failure limit parameter to obtain the gate wall corrosion rate under each typical operating condition parameter and failure limit parameter. Finally, a database is formed by each typical operating condition parameter and failure limit parameter and their corresponding gate wall corrosion rate.

(4) As shown in FIG2 , a neural network model based on the BP neural network algorithm is established. Then, the neural network is trained using the established database, and the training process is shown in FIG3 . The trained neural network can establish a multi-parameter mapping for each working condition. For example, the mapping of the acceleration grid aperture, upstream plasma density, atomic density, screen grid voltage and corrosion rate can be expressed as:

f(r _a ,n ₀ ,n _n ,U _s )→R _e (4)

Where, _ra , _n0 , _nn , _Us , and _Re represent the acceleration grid aperture, upstream plasma density, atomic density, screen grid voltage, and corrosion rate, respectively.

(5) For a single-mode working gate, the input parameters of the PIC-MCC model can be repeatedly called to continuously correct the input parameters of the final moment, and then the gate life performance parameters at the end of the life can be obtained by simulation calculation and the gate life can be given. Figure 4 shows the process of estimating the gate life performance parameters at the end of the life by iterative correction every 100 hours.

(6) For multi-mode wear and etching prediction, the essence of mode switching is the switching of mapping f, so the simulation scheme shown in Figure 5 is used for mode switching, that is, firstly obtain the mapping of different modes, and then use different mappings for different modes to iterate.

For the dynamic model of life prediction, this part requires that the final simulation tool can get rid of the use of PIC-MCC program for simulation calculation, and can quickly predict the gate life given any set of mode combinations. This part can be solved by adding gate failure criteria in the entire iterative process, and the specific process is shown in Figure 6.

Claims (6)

1. A method for dynamic prediction of ion thruster grid life based on artificial intelligence, characterized in that it includes the following steps: Step 1, establishing a gate simulation model based on a macro particle-Monte Carlo collision algorithm; Step 2, establishing a post-processing model for solving the gate corrosion rate, solving the corrosion rate distribution and average corrosion rate of the gate end surface and the hole wall; Step 3, forming a database from typical operating parameters, failure limit parameters and their corresponding gate wall corrosion rates; Estimate the gate failure limit parameters and calculate the accelerated gate electron reverse flow failure limit radius according to the following formula: Wherein, V _N is the net accelerating voltage, which is the sum of the screen grid voltage and the relative screen grid potential of the plasma in the ionization chamber; V _T is the total accelerating voltage, which is the difference between the net accelerating voltage and the accelerating grid voltage; l _g , _rs , _ta and _ra are the grid spacing, screen grid hole radius, accelerating grid thickness and accelerating grid hole radius respectively; Step 4, establishing a neural network model based on the BP neural network algorithm, using the database established in step 3 to train the neural network, and establishing multi-parameter mapping for different working conditions with the trained neural network; The trained neural network establishes multi-parameter mapping for different working conditions, including the mapping of acceleration grid aperture, upstream plasma density, atomic density, screen grid voltage and corrosion rate, which is expressed as: f(r _a ,n ₀ ,n _n ,U _s )→R _e (4) Where, _ra , _n0 , _nn , _Us , _Re represent the acceleration grid aperture, upstream plasma density, atomic density, screen grid voltage and corrosion rate respectively; Step 5: For the single-mode working gate, the input parameters based on the macro-particle-Monte Carlo collision model are continuously corrected by repeatedly calling the mapping to obtain the input parameters at the final moment, and then the gate life performance parameters at the end of the life are obtained by simulation calculation and the gate life is given; Step 6, in the multi-mode wear etching prediction, different mappings are used to iterate through different modes to achieve mode switching, and the final gate life is obtained by adding gate failure criteria in the entire iterative process. 2. The method for dynamic prediction of ion thruster grid life based on artificial intelligence according to claim 1 is characterized in that, in step 1, a grid simulation model based on a macroparticle-Monte Carlo collision algorithm is established, wherein the ratio of divalent ions to monovalent ions is calculated according to formula (1): In the formula, n _i ⁺⁺ , n _i ⁺ are the densities of divalent ions and monovalent ions respectively; I _d , I _b are the discharge current and beam current respectively; v ₀ , _vi are the atomic and ion velocities respectively; T _a , η _c are the gate transmittance and Clauin coefficient of atoms respectively; σ ₊₊ , σ ₊ are the divalent ion ionization collision cross sections and monovalent ion ionization collision cross sections respectively; γ is the ionization rate. 3. According to the method for dynamic prediction of ion thruster grid life based on artificial intelligence in claim 1, it is characterized in that in step 2, a post-processing program for solving the grid corrosion rate is established to solve the corrosion rate distribution and average corrosion rate of the grid end face and the hole wall, wherein the grid corrosion rate is calculated by formula (2): Where _RE is the corrosion rate, J is the gate wall ion current density, Y is the ion sputtering yield, q is the ion charge, _Mg is the atomic mass of the gate material, and _ρg is the gate material density. 4. According to the artificial intelligence-based dynamic prediction method for the life of an ion thruster grid, the characteristic is that, in step 3, the typical operating conditions and corresponding operating parameters include: acceleration grid thickness, acceleration grid aperture, deceleration grid thickness, deceleration grid aperture, screen grid-acceleration spacing, acceleration-deceleration spacing, screen grid voltage, acceleration grid voltage, deceleration grid voltage, upstream plasma density, neutral atom density, and anode voltage. 5. The method for dynamically predicting the life of an ion thruster grid based on artificial intelligence according to claim 1 is characterized in that, in step 5, the process of iteratively correcting and estimating the performance parameters of the grid at the end of its life is performed every 100 hours. 6. According to the artificial intelligence-based dynamic prediction method for the life of an ion thruster gate, according to claim 1, it is characterized in that, in step 6, in the multi-mode wear and etching estimation, different mappings are used in different modes to iterate and implement mode switching, and at the same time, the final life of the gate is obtained by adding a gate failure criterion in the entire iterative process, wherein the failure criterion is calculated by formula (3).