Plasma active group regulation and control method based on machine learning

Abstract

The invention relates to a plasma active group regulation and control method based on machine learning. The method comprises the following steps: acquiring plasma discharge condition data and plasma active group parameter data; calculating the contribution degree of the input features of each discharge condition to the target variable of the group property based on a gradient boosting tree model, obtaining a feature importance score, and screening out the input features with higher importance; an SHAP interpretation model is adopted, marginal contribution of each screened input feature to a target variable is measured through an SHAP value, and the input features which have remarkable influence on plasma active group parameter data are screened out; and establishing a nonlinear regression model of plasma discharge parameters and active group parameters, predicting the plasma active group parameters, and automatically adjusting the discharge parameters through a closed-loop feedback mechanism. Compared with the prior art, the method has the advantages that the analysis efficiency is remarkably improved, and the accuracy of predicting and regulating the properties of the plasma active groups under multi-parameter coupling is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of plasma active component regulation, and in particular to a plasma active group regulation method based on machine learning. Background Art

[0002] Low-temperature plasma is widely used in the biomedical field due to its characteristics of non-equilibrium and generation of diverse active groups, including disinfection and sterilization, cancer cell treatment and tissue repair. At present, scientists generally believe that the biological effects of plasma are mainly achieved through the active groups generated during plasma discharge. However, during the atmospheric pressure plasma discharge process, a variety of active groups (such as O, H2O2, ·OH, NO2 - The active group conditions (type, concentration and spatial distribution, etc.) are regulated by the plasma discharge parameters (such as gas conditions, driving power parameters, etc.).

[0003] Traditional methods for regulating plasma active components mostly use a step-by-step experimental method, that is, by changing a single parameter to analyze and study the changes in the properties of plasma active groups. This method is not only inefficient, but also difficult to reveal the synergistic effects between high-dimensional variables. In addition, there is usually a complex nonlinear coupling relationship between plasma discharge parameters. Traditional regulation methods are difficult to quickly screen the optimal parameter combination and cannot achieve accurate and efficient regulation of the properties of plasma active groups. For example, a plasma-activated water preparation integrated device and method disclosed in publication number CN118929852A achieves targeted regulation of the concentration of plasma active components by adjusting the electrode structure and pulse parameters. The problem of multivariable regulation of the properties of plasma active groups has always limited the further application of plasma technology in the biomedical field. Summary of the Invention

[0004] The purpose of the present invention is to provide a plasma active group control method based on machine learning in order to overcome the defects of the above-mentioned prior art that it is difficult to quickly screen the optimal parameter combination and cannot achieve accurate and efficient control of the properties of plasma active groups.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for regulating plasma active groups based on machine learning, comprising the following steps:

[0007] Collect plasma discharge condition data and plasma active group parameter data and perform preprocessing;

[0008] The input features are constructed based on the plasma discharge condition data, and the target variables are constructed based on the plasma active group parameter data. The contribution of each input feature to the target variable is calculated based on the gradient boosting tree model to obtain the feature importance score, and the input features with higher importance are screened out.

[0009] The SHAP interpretation model is used to measure the marginal contribution of each input feature to the target variable through the SHAP value, thereby screening out input features that have a significant impact on the plasma active group parameter data;

[0010] Based on the input features screened by the SHAP interpretation model, a nonlinear regression model of plasma discharge parameters and active group parameters was established;

[0011] Plasma active group parameters are predicted based on the nonlinear regression model, and plasma discharge parameters are automatically adjusted through a closed-loop feedback mechanism.

[0012] Furthermore, the plasma discharge condition data includes driving power supply parameters, gas conditions and ambient environment conditions. The driving power supply parameters include voltage waveform, amplitude, rising edge, falling edge, pulse width and frequency. The gas conditions include gas type, ratio and flow rate. The ambient environment conditions include temperature and humidity.

[0013] Furthermore, the plasma active group parameter data includes the type, concentration and spatial distribution of the plasma active groups.

[0014] Furthermore, the preprocessing includes processing missing values, data standardization and normalization.

[0015] Furthermore, the gradient boosting tree model uses split gain as an importance indicator to measure the contribution of each input feature to the target variable;

[0016] The split gain is used to reflect the cumulative contribution of each input feature to the reduction of the variance of the target variable when the model decision tree splits the node.

[0017] Furthermore, in the input feature screening process based on the gradient boosting tree model, each input feature is first sorted according to the score of the importance index to generate a feature importance ranking table;

[0018] Then, use a bar chart or heat map to display the importance ranking of input features according to the feature importance ranking table, to assist in identifying key control parameters for input feature screening.

[0019] Furthermore, the SHAP explanation model calculates the marginal contribution of each input feature to the target variable and analyzes the global explanation, local explanation, and interaction analysis of each input feature, thereby screening the input features.

[0020] The global explanation is used to generate a marginal contribution summary graph based on the marginal contribution of each input feature to the target variable, showing the positive or negative impact trend of each input feature on the target variable;

[0021] The local explanation is used to generate a marginal contribution force-directed graph for a single sample data to clarify the specific contribution of each input feature under the sample data;

[0022] The interaction analysis is used to calculate the interaction attribution value between two input features based on marginal contributions and visualize the marginal effects of key features through dependency graphs.

[0023] Furthermore, the method also includes evaluating the performance of the nonlinear regression model through mean square error, coefficient of determination and mean absolute percentage error.

[0024] Furthermore, the closed-loop feedback mechanism includes:

[0025] When the predicted value of the plasma active group parameter of the nonlinear regression model deviates from the corresponding target value, the plasma discharge parameters are automatically adjusted until the predicted value of the plasma active group parameter meets the target.

[0026] Furthermore, the plasma discharge parameters are provided with corresponding parameter adjustment thresholds.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] (1) The present invention analyzes the experimental data on plasma discharge parameters and active group properties. First, the contribution of each plasma discharge parameter to the active group properties is calculated using a gradient boosting tree model, thereby screening out plasma discharge parameter features with higher importance. On this basis, further screening is performed in combination with the SHAP interpretation model to measure the marginal contribution of the plasma discharge parameters to the active group properties, thereby obtaining plasma discharge parameters that have a significant impact on the active group properties and realizing rapid screening of the optimal plasma discharge parameters.

[0029] The nonlinear regression model of plasma discharge parameters and active group parameters constructed in this way can achieve accurate prediction and dynamic regulation of the conditions for the generation of plasma active groups.

[0030] (2) Experimental results show that compared with the traditional stepwise experimental method, this method significantly improves the analysis efficiency and effectively enhances the accuracy of predicting and controlling the properties of plasma-active groups under multi-parameter coupling. This invention can be widely used in the research and development of low-temperature plasma-related equipment in biomedical fields such as disinfection and sterilization, cancer treatment, etc., and has important scientific value and practical application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of a process for regulating plasma active groups based on machine learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0033] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0034] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0035] Example 1

[0036] like Figure 1 As shown, this embodiment provides a plasma active group control method based on machine learning, comprising the following steps:

[0037] S1: Collect plasma discharge condition data and plasma active group parameter data, and perform preprocessing;

[0038] S2: Construct input features based on plasma discharge condition data, construct target variables based on plasma active group parameter data, calculate the contribution of each input feature to the target variable based on the gradient boosting tree model, obtain feature importance scores, and screen out input features with higher importance;

[0039] S3: Using the SHAP explanatory model, the marginal contribution of each input feature after screening to the target variable is measured, thereby screening out input features that have a significant impact on the plasma active group parameter data;

[0040] S4: Based on the input features screened by the SHAP interpretation model, a nonlinear regression model of plasma discharge parameters and active group parameters is established;

[0041] S5: Predicting plasma active group parameters according to the nonlinear regression model, and automatically adjusting plasma discharge parameters through a closed-loop feedback mechanism.

[0042] Specifically, in step S1, the plasma discharge condition data includes driving power supply parameters, gas conditions and ambient environment conditions.

[0043] The driving power supply parameters include voltage waveform, amplitude, rising edge, falling edge, pulse width and frequency;

[0044] The gas conditions include gas type, ratio and flow rate;

[0045] The ambient environmental conditions include temperature and humidity.

[0046] Plasmon active group parameter data include the type, concentration and spatial distribution of plasma active groups.

[0047] Specifically, the process of collecting plasma active group parameter data includes:

[0048] By analyzing the types and relative concentrations of plasma active particles through absorption spectroscopy and emission spectroscopy, the parameters of the types, concentrations and spatial distribution of active groups under different conditions are obtained.

[0049] The types, absolute concentrations and spatial distribution parameters of plasma active particles are determined by laser induced fluorescence and Fourier transform infrared spectroscopy.

[0050] The preprocessing process in step S1 includes:

[0051] Process missing values, standardize and normalize the collected data;

[0052] The training data set is composed of plasma discharge condition data (gas flow, voltage, pulse frequency, etc.) and active group parameters (such as the concentration and spatial distribution of H2O2, ·OH, etc.).

[0053] Data format: Matrix form, each row represents an experimental sample, and the columns correspond to the input features (discharge parameters) and output targets (active group parameters).

[0054] In order to train the subsequent model, the data set is divided into a training set, a validation set, and a test set in a ratio of 6:2:2.

[0055] In step S2, the gradient boosting tree model uses split gain as an importance indicator to measure the contribution of each input feature to the target variable;

[0056] The split gain is used to reflect the cumulative contribution of each input feature to the reduction of the variance of the target variable when the model decision tree splits the node.

[0057] In the input feature screening process based on the gradient boosting tree model, first, the input features are sorted according to the scores of the importance indicators to generate a feature importance ranking table;

[0058] Then, use a bar chart or heat map to display the importance ranking of input features according to the feature importance ranking table, to assist in identifying key control parameters for input feature screening.

[0059] Specifically, based on the characteristics of the Gradient Boosting Tree (GBDT) model, the model's built-in "Feature Importance" function calculates the contribution of each input feature (such as gas flow, voltage, and pulse frequency) to the target variable (such as the concentration of reactive groups, such as H2O2). Specifically, "Split Gain" is used as an importance metric, reflecting the cumulative contribution of each feature to the reduction in variance of the target variable when the model's decision tree splits a node.

[0060] Output feature importance scores, sort them from high to low, and generate a feature importance ranking table (e.g., discharge voltage > gas flow > pulse frequency > ambient humidity).

[0061] Use bar charts or heat maps to visually display the feature importance ranking and assist in identifying key control parameters.

[0062] Combine domain knowledge to verify the rationality of the ranking results. For example, adjust the impact of high-importance parameters on active group parameters through experimental comparison to ensure that the physical meaning of the model is consistent with the actual situation.

[0063] Finally, key parameters were screened: sorting by importance score, retaining the top N key parameters (e.g., features with a total importance ≥ 80%), and eliminating redundant parameters (e.g., ambient humidity score < 5%).

[0064] In step S3, the SHAP (SHapley Additive exPlanations) explanation model is described as follows:

[0065] The SHAP explanation model calculates the marginal contribution of each input feature to the target variable and analyzes the global explanation, local explanation, and interaction analysis of each input feature to screen the input features.

[0066] The marginal contribution of each feature i to the predicted value f(x) is calculated as:

[0067]

[0068] It can be used to analyze the prediction results of a single sample and the interaction between features.

[0069] Global explanation is used to generate a marginal contribution summary chart based on the marginal contribution of each input feature to the target variable, showing the positive or negative impact trend of each input feature on the target variable;

[0070] That is, a SHAP summary graph (such as the Beeswarm graph) is generated to show the positive / negative influence trend of each feature on the predicted value in all samples, such as the positive correlation between the increase in discharge voltage and the H2O2 concentration, that is, voltage↑→H2O2↑, pulse frequency↑→·OH↓.

[0071] Local explanation is used to generate a marginal contribution force-directed graph for a single sample data to clarify the specific contribution of each input feature under the sample data;

[0072] That is, for a single experimental sample, a SHAP force plot is generated to clarify the specific contribution of each feature under the sample (for example: discharge voltage contribution +0.3, gas flow contribution -0.1).

[0073] Interaction analysis is used to calculate the interaction attribution value between two input features based on marginal contributions and visualize the marginal effects of key features through dependency plots.

[0074] Specifically, SHAP Interaction Values ​​are used to quantify the synergistic effects between features (e.g., the effect of the interaction between voltage and pulse frequency on ·OH concentration), and the marginal effects of key features are visualized through dependency plots.

[0075] Finally, combined with the SHAP analysis results, features that significantly affect the parameters of the target active groups (such as voltage and gas ratio) are screened to guide the priority adjustment direction of the experimental parameters.

[0076] That is, the optimal parameter combination generation process includes: for multiple active groups (such as H2O2 and ·OH), using the Pareto Front to balance different objectives, and giving priority to meeting the main requirements through SHAP value weight distribution.

[0077] Parameter space search: Combined with the SHAP analysis results, Bayesian optimization or genetic algorithm is used to search for the optimal combination within a reasonable range of key parameters (such as voltage and flow).

[0078] Based on the features extracted in the above steps, a LASSO regression model is established and the parameters are trained to obtain the final prediction model;

[0079] The data collection and modeling process in this embodiment is as follows:

[0080] 1. Use a normal-pressure dielectric barrier discharge (DBD) device to conduct plasma discharge experiments by adjusting the gas flow rate (1-5 L / min), discharge voltage (10-30 kV), and pulse frequency (1-5 kHz).

[0081] 2. Measure the concentration and spatial distribution of H2O2, ·OH, etc. through emission spectroscopy and other analysis to form an experimental data set.

[0082] 3. Use the GBDT algorithm to process the data and construct a nonlinear regression model of active group parameters and discharge parameters.

[0083] Evaluate the model and deploy it to the system for online prediction and dynamic regulation. Specifically, it includes:

[0084] (a) Model performance evaluation

[0085] The following metrics are used to evaluate the performance of the regression model:

[0086] Mean squared error (MSE): measures the average deviation between the predicted value and the experimental value;

[0087] Coefficient of determination (R 2 ): Quantify the model's ability to explain the variance of the target variable (target: R 2 ≥0.85);

[0088] Mean absolute percentage error (MAPE): assesses the relative error of the predicted active group parameters (target: MAPE ≤ 10%).

[0089] Verify the model's generalization ability on the test set to ensure stable indicators.

[0090] (b) Model deployment and integration

[0091] Convert the trained GBDT model into a lightweight format (such as ONNX or PMML) and embed it into the control system of the plasma equipment.

[0092] A real-time data interface is designed to collect discharge parameters (such as voltage, gas flow, etc.) in real time and input them into the model to output the predicted active group parameters.

[0093] (c) Dynamic control strategy

[0094] Build a closed-loop feedback mechanism:

[0095] When the predicted active group parameters deviate from the target values, the system automatically adjusts the discharge parameters (such as fine-tuning the voltage by ±1kV or the flow rate by ±0.5L / min) and re-predicts until the target is met.

[0096] Update the model regularly: Incrementally train the model based on new experimental data to improve prediction accuracy.

[0097] (d) Safety and redundancy design

[0098] Set parameter adjustment thresholds (such as voltage not exceeding the device rating) to prevent overload;

[0099] Deploy backup models (such as random forests) as redundancy to ensure system robustness.

[0100] Preferably, the above method further comprises: for the target active group parameter, reversely solving the SHAP value:

[0101] If the SHAP value of the current voltage is a positive contribution, the voltage is further increased to the device safety threshold;

[0102] If the SHAP value of the gas flow rate is a negative contribution, reduce the flow rate to optimize the active group parameters.

[0103] The control and verification process in this embodiment is as follows:

[0104] 1. Use the regression model to predict the optimal discharge parameter combination (gas flow rate 2.5L / min, discharge voltage 18kV, pulse frequency 3kHz).

[0105] 2. Under the predicted conditions, the H2O2 concentration increased by 10%-15% compared with the initial conditions, and further experiments verified that its microbial inactivation effect was significantly enhanced.

[0106] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A plasma active group control method based on machine learning, characterized in that: The following steps are involved: Collect plasma discharge condition data and plasma active group parameter data and perform preprocessing; The input features are constructed based on the plasma discharge condition data, and the target variables are constructed based on the plasma active group parameter data. The contribution of each input feature to the target variable is calculated based on the gradient boosting tree model to obtain the feature importance score, and the input features with higher importance are screened out. The SHAP explanatory model is used to measure the marginal contribution of each input feature to the target variable after screening, thereby screening out input features that have a significant impact on the plasma active group parameter data; Based on the input features screened by the SHAP interpretation model, a nonlinear regression model of plasma discharge parameters and active group parameters was established; Plasma active group parameters are predicted based on the nonlinear regression model, and plasma discharge parameters are automatically adjusted through a closed-loop feedback mechanism.

2. The plasma active group control method based on machine learning according to claim 1, characterized in that: The plasma discharge condition data includes driving power supply parameters, gas conditions and ambient environment conditions. The driving power supply parameters include voltage waveform, amplitude, rising edge, falling edge, pulse width and frequency. The gas conditions include gas type, ratio and flow rate. The ambient environment conditions include temperature and humidity.

3. The plasma active group control method based on machine learning according to claim 1, characterized in that: The plasma active group parameter data includes the type, concentration and spatial distribution of the plasma active groups.

4. The plasma active group control method based on machine learning according to claim 1, characterized in that: The preprocessing includes processing missing values, data standardization and normalization.

5. The plasma active group control method based on machine learning according to claim 1, characterized in that: The gradient boosting tree model uses split gain as an importance indicator to measure the contribution of each input feature to the target variable; The split gain is used to reflect the cumulative contribution of each input feature to the reduction of the variance of the target variable when the model decision tree splits the node.

6. The plasma active group control method based on machine learning according to claim 5, characterized in that: In the input feature screening process based on the gradient boosting tree model, first, the input features are sorted according to the scores of the importance indicators to generate a feature importance ranking table; Then, use a bar chart or heat map to display the importance ranking of input features according to the feature importance ranking table, to assist in identifying key control parameters for input feature screening.

7. The plasma active group control method based on machine learning according to claim 1, characterized in that: The SHAP explanation model calculates the marginal contribution of each input feature to the target variable and analyzes the global explanation, local explanation, and interaction analysis of each input feature to screen the input features. The global explanation is used to generate a marginal contribution summary graph based on the marginal contribution of each input feature to the target variable, showing the positive or negative impact trend of each input feature on the target variable; The local explanation is used to generate a marginal contribution force-directed graph for a single sample data to clarify the specific contribution of each input feature under the sample data; The interaction analysis is used to calculate the interaction attribution value between two input features based on marginal contributions and visualize the marginal effects of key features through dependency graphs.

8. The plasma active group control method based on machine learning according to claim 1, characterized in that: The method also includes evaluating the performance of the nonlinear regression model using mean squared error, coefficient of determination, and mean absolute percentage error.

9. The plasma active group control method based on machine learning according to claim 1, characterized in that: The closed-loop feedback mechanism includes: When the predicted value of the plasma active group parameter of the nonlinear regression model deviates from the corresponding target value, the plasma discharge parameters are automatically adjusted until the predicted value of the plasma active group parameter meets the target.

10. The plasma active group control method based on machine learning according to claim 9, characterized in that: The plasma discharge parameters are provided with corresponding parameter adjustment thresholds; The method further includes reversely solving the marginal contribution value of each plasma discharge condition parameter for the target value of the plasma active group parameter. If the marginal contribution is positive, the corresponding plasma discharge condition parameter is increased; otherwise, the corresponding plasma discharge condition parameter is reduced.

Images