Plasma air purifier control method and system based on artificial intelligence

Abstract

The invention discloses a plasma air purifier control method and system based on artificial intelligence, and the method comprises the steps: collecting the internal and surrounding environment data of a plasma air purifier in real time, extracting pollution space distribution characteristics through combining with an improved Point Net + + network, and constructing a knowledge graph containing the pollution concentration, equipment parameters and other multi-dimensional information. A graph neural network (GCN) is innovatively adopted to establish a nonlinear mapping model of working parameters and filtering efficiency, key parameter combinations such as voltage and wind speed are dynamically optimized through a strategy gradient reinforcement learning algorithm, and a strategy library with the maximum filtering efficiency is formed. And a multi-physics field coupling simulation technology is introduced to verify the reliability of the strategy, a heterogeneous knowledge graph is constructed, and rapid generation and accurate execution of a control instruction are realized. The difference between real data and twin data is compared in real time through a digital twin technology, a reinforcement learning strategy is dynamically adjusted, and a closed-loop optimization system is formed. The purification efficiency and the energy efficiency ratio of the plasma air purifier are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of air purification, and specifically to a plasma air purifier control method and system based on artificial intelligence. Background Art

[0002] Efficient air purification equipment has become the key to ensuring indoor air quality. Plasma air purifiers occupy an important position in the field of air purification due to their unique purification principle. It generates plasma by ionizing air, and uses the active ingredients of plasma to decompose and remove pollutants in the air, and has a good purification effect on a variety of pollutants.

[0003] However, the current control technology of plasma air purifiers has many bottlenecks, which greatly limits its performance.

[0004] First, when the air purifier is working, the discharge voltage of the needle electrode array, the electrostatic field strength between the metal foam and the grounded mesh electrode, the wind speed in the ventilation duct and other parameters contain rich information, which are interrelated and jointly affect the purification effect. However, the traditional method simply processes these data, fails to fully explore the potential laws and values ​​behind the data, and cannot establish an accurate mathematical relationship model between parameters and purification efficiency, resulting in a lack of strong data support when optimizing purifier performance.

[0005] Secondly, indoor pollution distribution is not uniform, and the pollution degree and pollutant types in different areas may vary greatly. However, traditional purifiers often adopt a fixed working mode and cannot perform targeted purification according to the spatial changes of pollution. They are insufficient in severely polluted areas and cause energy waste in less polluted areas, resulting in low overall purification efficiency and difficulty in meeting the complex and changing pollution control needs of indoor areas.

[0006] Thirdly, the indoor environment is changing dynamically. The entry and exit of people, the opening and closing of doors and windows, the generation of new pollutants, etc. will change the indoor pollution situation. However, traditional purifiers lack the ability to perceive environmental changes in real time and adjust the working state accordingly. They cannot optimize the working parameters in time according to the actual pollution situation, resulting in unstable purification effect and failure to provide users with a consistent high-quality air environment.

[0007] Traditional methods also have limitations in terms of optimization strategies and technical means. Traditional plasma air purifiers lack a systematic optimization method when determining working parameters, making it difficult to find a parameter combination that maximizes filtration efficiency and minimizes energy consumption. At the same time, traditional technologies rarely use multi-physics field coupling simulation technology to conduct in-depth analysis of the complex physical processes inside the purifier, and are unable to accurately predict the purification effect under different working conditions, resulting in a lack of scientific basis for improving the performance of the purifier.

[0008] To sum up, the existing plasma air purifier control technology cannot fully exert its purification potential. There is an urgent need for an innovative control method and system that can accurately sense the pollution status, deeply analyze the data, and realize intelligent adaptive control. With the help of advanced simulation and optimization technology, the performance of plasma air purifiers can be comprehensively improved to create a healthier and more comfortable indoor air environment for people. Summary of the invention

[0009] The technical problem to be solved by the present invention is to provide a plasma air purifier control method and system based on artificial intelligence, aiming to solve the problems of inaccurate pollution perception, low parameter optimization efficiency, and insufficient adaptive ability in traditional control technology.

[0010] To solve the above technical problems, an embodiment of the present invention provides the following technical solution: a plasma air purifier control method based on artificial intelligence, comprising the following steps:

[0011] The working parameters of the plasma air purifier system were collected, and the pollution spatial distribution characteristics in the plasma air purifier were extracted through the improved PointNet++ network to build a pollution feature knowledge base.

[0012] The pollution feature knowledge base is connected to the graph neural network model, and the graph neural network model is used to model and train the collected working parameters and the system's filtration efficiency for aerosols to obtain a graph neural network model that can accurately predict filtration efficiency.

[0013] The reinforcement learning algorithm is combined with the trained graph neural network model to optimize the system's working parameters and form a strategy library for the working parameter combination that maximizes filtering efficiency.

[0014] Use multi-physics field coupling simulation software to simulate and analyze the plasma air purification system, associate the simulation results with the reinforcement learning strategy library, and generate a parameter optimization knowledge graph;

[0015] Generate plasma air purifier control instructions based on the knowledge graph to control the working state of the plasma air purifier;

[0016] Iteratively detect the spatial distribution of pollution, and dynamically correct the reinforcement learning strategy through difference analysis between twin data and real data.

[0017] Furthermore, the improved PointNet++ network is used to extract the pollution spatial distribution characteristics in the plasma air purifier, specifically:

[0018] Use sensors to collect spatial distribution data of pollution in and around the plasma air purifier, convert it into point cloud data containing location and pollution information, and normalize each feature dimension of the data;

[0019] The improved PointNet++ network structure is used for hierarchical sampling, feature extraction and feature propagation. The sampling layer uses the farthest point sampling to select key points from the current layer point cloud, and the grouping layer finds the nearest K points for each key point in the current layer point cloud to form a local neighborhood. Then the PointNet module is used to extract local neighborhood features, and the feature values ​​are calculated for the current layer points through inverse distance weighted interpolation. The interpolated features are concatenated with the original features of the current layer, and the new feature matrix is ​​obtained through MLP processing.

[0020] After multi-layer processing, the network outputs the global feature vector of each point, and then obtains the global features of the contaminated space through average pooling or maximum pooling.

[0021] Furthermore, the construction of the pollution feature knowledge base specifically includes:

[0022] Preprocess the collected working parameters for data cleaning and data standardization;

[0023] The principal component analysis was used to reduce the dimension and extract the principal component. The clustering algorithm K-Means was used to calculate the distance between the sample and the cluster center and continuously adjust the center to achieve clustering, discover the distribution pattern of pollution data, and distinguish areas with different pollution levels.

[0024] The graph database Neo4j is used to store the extracted features of the working parameters and the pollution space distribution features. Nodes representing entities and edges representing entity relationships are created in the graph database, and data are collected and updated periodically to complete data construction.

[0025] Furthermore, the pollution feature knowledge base is connected to the graph neural network model, and the graph neural network model is used to model and train the collected working parameters and the system's aerosol filtration efficiency, specifically:

[0026] Extract data from the pollution feature knowledge base, normalize numerical data and perform one-hot encoding on categorical data during preprocessing;

[0027] Taking pollution characteristics and working parameters as nodes, edges are built according to the association relationship, and the adjacency matrix A is constructed accordingly. If there is an edge between node i and node j, then A ij =1, otherwise A ij =0; at the same time, the preprocessed data is combined into a feature matrix X to complete the construction of the graph data structure;

[0028] Select the graph neural network model GCN, the calculation formula of the GCN layer is:

[0029]

[0030] Among them, H (l) is the node feature matrix of the lth layer, W (l)is the weight matrix of the lth layer, σ is the activation function, is the adjacency matrix with self-connection added, I is the identity matrix, yes The degree matrix, whose elements

[0031] By stacking multiple GCN layers, the model can learn more complex graph structure features. After the last layer of GCN, a fully connected layer is added as the output layer to map the node features to the predicted filtering efficiency. Assume that the node feature matrix after multi-layer GCN processing is H (L) , the calculation formula of the output layer is:

[0032] y=FC(H (L) )

[0033] Where y is the predicted filtering efficiency, FC represents the fully connected layer operation;

[0034] The mean square error (MSE) loss function is used to measure the difference between the filtration efficiency predicted by the model and the actual filtration efficiency. The loss function formula is:

[0035]

[0036] Where n is the number of samples, y i is the actual filtration efficiency, is the filtration efficiency predicted by the model.

[0037] Input the constructed graph data into the model, calculate the predicted value of the model through forward propagation, then calculate the loss value according to the loss function, use the optimizer to update the model parameters through back propagation, and continuously iterate the training until the loss value converges or reaches the preset number of training rounds.

[0038] Furthermore, the reinforcement learning algorithm is combined with the trained graph neural network model to optimize the working parameters of the system to form a strategy library of working parameter combinations that maximize filtering efficiency, specifically:

[0039] The operating environment of the plasma air purifier is modeled as a Markov decision process, the state space, action space and reward function are defined, and GNN is used to perform graph structure modeling on the topological relationship between system parameters and pollution feature knowledge base;

[0040] Construct a GNN-based policy network Actor and a value function network Critic to dynamically generate parameter optimization strategies;

[0041] The policy gradient algorithm is used to update the policy network to maximize the long-term cumulative reward.

[0042] Furthermore, the multi-physics field coupling simulation software is used to simulate and analyze the plasma air purification system, and the simulation results are associated with the reinforcement learning strategy library to generate a parameter optimization knowledge graph, specifically:

[0043] (1) Establish an electric field-flow field-concentration field coupling simulation model of the plasma purification system, including the electric field control equation, flow field control equation, and pollutant control equation, where:

[0044] The governing equation for the electric field is: φ is the electric potential, ε is the dielectric constant, and ρ is the space charge density;

[0045] The governing equations of the flow field are:

[0046]

[0047] u describes the velocity of a point in the fluid at a certain moment, including the magnitude and direction of the velocity; t indicates the order of occurrence and duration of the physical process; p is the force acting vertically on the unit area of ​​the fluid; u is the dynamic viscosity of the fluid; F e is the electric force;

[0048] Pollutant transport equation:

[0049]

[0050] C refers to the concentration of pollutants in space, that is, the content of pollutants per unit volume; D is the diffusion coefficient. The larger the diffusion coefficient, the faster the pollutants diffuse; k represents the speed of the reaction between plasma and pollutants;

[0051] Define the simulation input parameter x sim =[V, u, d, T, ε], V, u, d, T, ε represent voltage, flow rate, electrode spacing, temperature, medium properties, and output performance index y sim =[η, P, ΔC], η, P, ΔC represent filtration efficiency, energy consumption, and concentration gradient, respectively;

[0052] (5) Associating simulation data with the policy library, mapping the parameter combination in the reinforcement learning policy library to the simulation input, and calculating the difference between the simulation result and the expected performance of the policy library. If the difference is less than a threshold, the parameter combination is marked as a credible policy, otherwise simulation calibration is triggered;

[0053] (6) Construct a heterogeneous knowledge graph, which contains three types of nodes and edges, where the nodes include parameter nodes (x sim =[V,u,d,T,ε]), performance node (y sim =[η, P, ΔC]), strategy node (π 1 , π 2 ,…π n), edge types include causal relationship, policy association, and simulation verification;

[0054] (7) Generate node embedding using graph attention network GAT

[0055]

[0056] represents the embedding representation of node i in the l+1th layer of the graph attention network; σ is the activation function; It means to sum the neighbor node set N(i) of node i, where N(i) includes all nodes directly connected to node i;

[0057] α ij is the attention coefficient, which is used to measure the degree of association between node i and its neighbor node j; W (l) is the weight matrix of the lth layer; represents the embedding representation of neighbor node j at layer l.

[0058] Furthermore, the plasma air purifier control instructions are generated according to the knowledge graph to control the working state of the plasma air purifier, specifically:

[0059] Acquire real-time data of the environment and equipment through sensors and map them into state vectors that can be recognized by the knowledge graph;

[0060] Based on the state vector, the optimal control strategy matching the current state is retrieved in the knowledge graph through the cosine similarity matching method;

[0061] Convert policy parameters into physical control signals executable by the device;

[0062] Sending instructions to the plasma generator execution unit through the industrial bus protocol;

[0063] Combined with sensor feedback, control parameters are corrected in real time to form a closed-loop optimization.

[0064] Furthermore, the strategy parameters are converted into physical control signals executable by the device, specifically using a PID control algorithm to control the voltage of the plasma generator by adjusting the pulse width modulation PWM signal. The calculation formula is:

[0065] PWM v =K p ·(V opt -V real )+K i ·∫(V opt -V real )dt

[0066] Among them, PWM vK is a pulse width modulation (PWM) signal used to control the voltage of the plasma generator; p is the proportionality coefficient, K i is the integral coefficient; V opt is the optimal voltage value, V real It is the actual measured voltage value of the plasma generator.

[0067] Furthermore, the iterative detection of the spatial distribution of pollution dynamically corrects the reinforcement learning strategy through the difference analysis between the twin data and the real data, specifically:

[0068] Use sensors to continuously collect real data on the spatial distribution of pollution in and around the plasma air purifier;

[0069] Based on the established pollution feature knowledge base, system model and previously accumulated data, digital twin technology is used to generate twin data corresponding to the real scene;

[0070] The collected real data and generated twin data are normalized, and the mean square error Euclidean distance indicator is used to calculate the difference between the twin data and the real data;

[0071] According to the dimension of significant difference, the reward function in reinforcement learning is adjusted. Assuming that the original reward function is R(s,a), the adjustment factor based on the difference is α, and the adjusted reward function is R`(s,a), then:

[0072] R`(s,a)=αR(s,a), where a is the action and s is the state;

[0073] Using reinforcement learning algorithms such as the policy gradient algorithm, the policy network is updated using the adjusted reward function. The update formula for the policy network parameter θ is:

[0074]

[0075] Among them, θ t is the parameter of the current policy network, η is the learning rate, π θ (a|s) is the probability of taking action a in state s;

[0076] Integrate the newly collected real data and calculated difference information into the pollution characteristic knowledge base, and update and optimize the pollution data distribution pattern and working parameter characteristics;

[0077] Based on the results of strategy correction, the parameters of the graph neural network model, reinforcement learning model, and simulation model are fine-tuned.

[0078] The present invention also proposes a plasma air purifier control system based on artificial intelligence, comprising:

[0079] The pollution feature knowledge base construction module collects the working parameters of the plasma air purifier system, extracts the pollution spatial distribution characteristics in the plasma air purifier through the improved PointNet++ network, and constructs the pollution feature knowledge base;

[0080] A graph neural network model building module is used to connect the pollution feature knowledge base to the graph neural network model, use the graph neural network model to model and train the collected working parameters and the system's filtration efficiency for aerosols, and obtain a graph neural network model that can accurately predict filtration efficiency;

[0081] The strategy library building module is used to optimize the system's working parameters by using a reinforcement learning algorithm combined with a trained graph neural network model to form a strategy library of working parameter combinations that maximize filtering efficiency;

[0082] A knowledge graph construction module is used to use multi-physics field coupling simulation software to simulate and analyze the plasma air purification system, associate the simulation results with the reinforcement learning strategy library, and generate a parameter optimization knowledge graph;

[0083] A control instruction generation module is used to generate plasma air purifier control instructions according to the knowledge graph to control the working state of the plasma air purifier;

[0084] The strategy optimization module is used to iteratively detect the spatial distribution of pollution and dynamically correct the reinforcement learning strategy through difference analysis between twin data and real data.

[0085] The beneficial effects of the above technical solution of the present invention are as follows:

[0086] 1. This invention realizes the three-dimensional feature extraction and real-time analysis of pollution spatial distribution through the improved PointNet++ network and multimodal sensor fusion technology. This method breaks through the limitations of traditional single-point monitoring, can accurately identify indoor pollution hotspots, and dynamically adjust purification strategies according to pollution concentrations.

[0087] 2. The present invention combines graph neural networks with reinforcement learning algorithms to establish a nonlinear mapping model between working parameters and purification efficiency. The strategy gradient algorithm optimizes the combination of key parameters such as voltage and wind speed, effectively reducing energy consumption while ensuring efficient purification. It can adapt to different pollution scenarios and dynamically generate optimal control strategies, so that the purifier always maintains the best operating state in complex environments, achieving dual optimization of purification performance and energy efficiency ratio.

[0088] 3. The present invention constructs a multi-physics field coupling simulation and digital twin verification system, and verifies the parameter combination generated by reinforcement learning through virtual environment simulation. When the simulation result deviates from the actual effect, the model calibration is automatically triggered to ensure the reliability of the strategy. This mechanism reduces the need for real machine testing, reduces the risk of equipment loss, provides a scientific basis for the optimization strategy, and ensures long-term stable operation.

[0089] 4. The present invention establishes a dynamic correction mechanism for reinforcement learning strategies through differential analysis between twin data and real data. When pollution distribution or environmental conditions change, the system can quickly adjust working parameters. Combining PID control algorithm with industrial bus protocol to transmit control instructions to the purifier, a closed-loop optimization system is formed to ensure that the purifier can maintain efficient and stable operation under complex working conditions such as pollution mutations.

[0090] 5. The present invention constructs a heterogeneous knowledge graph of parameters, performance and strategies, integrates multi-source data and optimization strategies, and realizes rapid retrieval of real-time status and optimal strategies through cosine similarity matching method, accelerating the generation of control instructions. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 This is a flow chart of the plasma air purifier control method based on artificial intelligence of the present invention;

[0092] Figure 2 This is a principle block diagram of the plasma air purifier control system based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0093] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0094] like Figure 1 As shown, the present invention proposes a plasma air purifier control method based on artificial intelligence, comprising the following steps:

[0095] S1. Collect the working parameters of the plasma air purifier system, extract the pollution spatial distribution characteristics in the plasma air purifier through the improved PointNet++ network, and build a pollution feature knowledge base;

[0096] S2. Connect the pollution feature knowledge base to the graph neural network model, use the graph neural network model to model and train the collected working parameters and the system's filtration efficiency for aerosols, and obtain a graph neural network model that can accurately predict filtration efficiency;

[0097] S3. Use reinforcement learning algorithm combined with the trained graph neural network model to optimize the system's working parameters and form a strategy library of working parameter combinations that maximize filtering efficiency.

[0098] S4. Use multi-physics field coupling simulation software to simulate and analyze the plasma air purification system, associate the simulation results with the reinforcement learning strategy library, and generate a parameter optimization knowledge graph;

[0099] S5. Generate a plasma air purifier control instruction according to the knowledge graph to control the working state of the plasma air purifier;

[0100] S6. Iteratively detect the spatial distribution of pollution and dynamically correct the reinforcement learning strategy through difference analysis between twin data and real data.

[0101] In this embodiment, the improved PointNet++ network is used to extract the pollution spatial distribution characteristics in the plasma air purifier, specifically:

[0102] S11. Use sensors to collect spatial distribution data of pollution in and around the plasma air purifier, convert it into point cloud data containing location and pollution information, and normalize each feature dimension of the data. Among them, a combination of multimodal sensors (such as laser dust sensors, temperature and humidity sensors, VOC sensors) is used, distributed in a grid pattern inside and around the purifier, and a micro laser radar (LiDAR) is deployed to obtain three-dimensional spatial coordinate information with an accuracy of up to ±2mm. The sensor sampling frequency is set to 10Hz to ensure real-time capture of dynamic pollution changes. When generating point cloud data, the sensor position is mapped to a three-dimensional Cartesian coordinate system (X, Y, Z), with the origin set at the center of the purifier, and spatial coordinate encoding is performed. Each point contains 5-dimensional features (PM2.5, PM10, VOC, temperature and humidity, CO 2 ), and then the pollution features are fused. The point cloud is stored in PLY (Polygon File Format) format to support subsequent deep learning processing.

[0103] S12. Use the improved PointNet++ network structure for hierarchical sampling, feature extraction and feature propagation; the sampling layer uses the farthest point sampling to select key points from the current layer point cloud, and the grouping layer finds the nearest K points in the current layer point cloud for each key point to form a local neighborhood; then use the PointNet module to extract local neighborhood features, calculate feature values ​​for the current layer points through inverse distance weighted interpolation, splice the interpolated features with the original features of the current layer, and obtain a new feature matrix through MLP processing. The farthest point sampling method can make the sampling points more evenly distributed in the point cloud space and avoid the sampling points from gathering in local areas. For example, in a plasma air purifier, it can cover different pollution concentration areas more comprehensively to ensure that various pollution features can be collected. The grouping layer selects the nearest K points to form a local neighborhood. This K value will affect the fineness of the local features. The larger the K value, the richer the local information, but the amount of calculation will also increase. It is generally determined according to the actual data characteristics and computing resources. The PointNet module extracts local neighborhood features and can capture the unique characteristics of each local area, such as local aggregation and discreteness of pollutant distribution. Inverse distance weighted interpolation combines the information of surrounding points to calculate the feature value of the current point. The closer the point is, the greater the impact on the current point is, which can more accurately reflect the local change trend. The interpolated features are spliced ​​with the original features to integrate feature information from different sources and enrich the feature expression. After multi-layer perceptron (MLP) processing, the complex relationship between features can be further explored, and a new feature matrix can be generated to provide more valuable data for subsequent analysis.

[0104] S13. After multi-layer processing, the network outputs the global feature vector of each point, and then obtains the global features of the contaminated space through average pooling or maximum pooling. After multi-layer processing, the network outputs the global feature vector of each point, which contains the local and global information of each point in the point cloud data. However, since point cloud data is usually disordered and the number of points is large, directly using the feature vectors of these points for subsequent analysis may lead to problems of high computational complexity and information redundancy. Therefore, through average pooling or maximum pooling operations, the feature vectors of these points can be aggregated into a single global feature vector, thereby reducing the data dimension, retaining the most important information, and improving the generalization ability of the model.

[0105] In this embodiment, the step S1 constructs a pollution feature knowledge base, specifically including:

[0106] S141. Preprocess the collected working parameters for data cleaning and data standardization. Data cleaning includes missing value processing, outlier detection, and duplicate value processing. Data standardization includes Min-Max normalization, etc.

[0107] S142, use principal component analysis to reduce the dimension and extract the principal component, use the clustering algorithm K-Means to calculate the distance between the sample and the cluster center and continuously adjust the center to achieve clustering, find the distribution pattern of polluted data, and distinguish areas with different pollution levels. Principal component analysis is a commonly used dimensionality reduction technology. Its core principle is to transform multiple related variables in the original data into a few unrelated comprehensive variables, namely principal components, through linear transformation. The K-Means algorithm is a distance-based clustering algorithm, the purpose of which is to divide the data into K clusters, so that the data points in the same cluster have a high similarity, while the data points between different clusters have a low similarity. In the processing of polluted data, it randomly selects K initial cluster centers, and then calculates the distance from each sample point to each cluster center, and assigns the sample point to the cluster with the nearest cluster center. Then, recalculate the cluster center of each cluster, that is, the mean of all sample points in the cluster. Repeat this process until the cluster center no longer changes or changes very little, that is, reaches a convergence state. Effectively reduce the data dimension through principal component analysis, and combine the optimized K-Means algorithm to achieve intelligent division of polluted areas.

[0108] S143. Use the graph database Neo4j to store the extracted features of the working parameters and the pollution space distribution features, create nodes representing entities and edges representing entity relationships in the graph database, and collect and update them periodically to complete data construction.

[0109] Node design includes the design of working parameter nodes and pollution spatial distribution characteristic nodes, among which:

[0110] Working parameter nodes: Create nodes for each working parameter, such as discharge voltage node, electrostatic field strength node, wind speed node, etc. Each node has specific attributes. For example, the discharge voltage node contains attributes such as voltage value, measurement time, and measurement location. These attributes can accurately describe the specific information of the working parameters, which is convenient for subsequent query and analysis.

[0111] Pollution spatial distribution feature nodes: Create nodes based on pollution type (such as PM2.5, VOC, etc.) and spatial location. For example, a node representing the PM2.5 concentration at a certain spatial location contains attributes such as PM2.5 concentration value, measurement time, and spatial coordinates. Through these attributes, you can clearly understand the distribution of pollution in different spaces and times.

[0112] The edge design includes causal edges, spatiotemporal association edges, and device association edges, among which:

[0113] Causal edge: used to represent the impact of operating parameters on pollution characteristics. For example, an edge is created from the discharge voltage node to the PM2.5 concentration node. The attributes of the edge can include the degree of influence (such as positive correlation or negative correlation) and the influence coefficient, etc., to quantify the impact of operating parameters on pollution characteristics.

[0114] Spatiotemporal correlation edge: represents the spatial and temporal correlation between different pollution feature nodes. For example, edges are created between PM2.5 concentration nodes at adjacent spatial locations in the same time period. The attributes of the edges can contain information such as spatial distance and time interval, which helps to analyze the propagation patterns of pollution in space and time.

[0115] Device association edge: connects the working parameter node and the corresponding purifier device node, indicating which purifier generates the working parameter, which facilitates device management and troubleshooting.

[0116] In this embodiment, the step S2 connects the pollution characteristic knowledge base to the graph neural network model, and uses the graph neural network model to model and train the collected working parameters and the system's aerosol filtration efficiency. Traditional machine learning models often find it difficult to capture high-order relationships and global information between nodes when processing complex graph structure data. The graph neural network (GNN) can adapt well to graph structure data. Through the message passing mechanism, nodes can aggregate the information of their neighbor nodes, thereby learning the feature representation of nodes and edges. When the pollution characteristic knowledge base is connected to the graph neural network model, GNN can effectively handle the complex correlation between working parameters, pollution characteristics and filtration efficiency, and mine the potential patterns in the data, providing more accurate results for modeling and predicting aerosol filtration efficiency. The specific steps are:

[0117] S21, extract data from the pollution feature knowledge base, normalize the numerical data during preprocessing, and perform one-hot encoding on the categorical data;

[0118] S22, using pollution characteristics and working parameters as nodes, building edges based on the association relationship, and constructing the adjacency matrix A accordingly. If there is an edge between node i and node j, then A ij =1, otherwise A ij =0; at the same time, the preprocessed data is combined into a feature matrix X to complete the construction of the graph data structure;

[0119] Select the graph neural network model GCN, the calculation formula of the GCN layer is:

[0120]

[0121] Among them, H (l) is the node feature matrix of the lth layer, W (l) is the weight matrix of the lth layer, σ is the activation function, is the adjacency matrix with self-connection added, I is the identity matrix, yes The degree matrix, whose elements

[0122] S23. By stacking multiple GCN layers, the model can learn more complex graph structure features. After the last layer of GCN, add a fully connected layer as the output layer to map the node features to the predicted filtering efficiency. Assume that the node feature matrix after multi-layer GCN processing is H (L) , the calculation formula of the output layer is:

[0123] y=FC(H (L) )

[0124] Where y is the predicted filtering efficiency, FC represents the fully connected layer operation;

[0125] The mean square error (MSE) loss function is used to measure the difference between the filtration efficiency predicted by the model and the actual filtration efficiency. The loss function formula is:

[0126]

[0127] Where n is the number of samples, y i is the actual filtration efficiency, is the filtration efficiency predicted by the model.

[0128] S24. Input the constructed graph data into the model, calculate the predicted value of the model through forward propagation, then calculate the loss value according to the loss function, use the optimizer to update the parameters of the model through back propagation, and continuously iterate the training until the loss value converges or reaches the preset number of training rounds.

[0129] In this embodiment, step S3 uses a reinforcement learning algorithm combined with a trained graph neural network model to optimize the working parameters of the system to form a strategy library of working parameter combinations that maximize filtering efficiency, specifically:

[0130] S31. Model the operating environment of the plasma air purifier as a Markov decision process, define the state space, action space and reward function, and use GNN to perform graph structure modeling on the topological relationship between system parameters and pollution feature knowledge base;

[0131] State space definition: The state space contains key information when the plasma air purifier is running, such as operating parameters (discharge voltage of the needle electrode array, electrostatic field strength between the metal foam and the grounded mesh electrode, wind speed in the ventilation duct, distance between the needle electrode array and the metal foam, and PPI of the metal foam), as well as pollution spatial distribution characteristics. This information comprehensively describes the current operating state of the purifier and provides a basis for subsequent decision-making.

[0132] Action space definition: Action space refers to the various operations that the purifier can take, specifically the adjustment of working parameters. For example, changing the discharge voltage, adjusting the wind speed, etc. These actions will affect the operating status and filtration efficiency of the purifier.

[0133] Reward function definition: The reward function is the key to guiding the optimization of the purifier. When setting the reward function, the improvement of filtration efficiency is used as a positive reward, and the increase in energy consumption is used as a negative reward. If the purifier improves the filtration efficiency without significantly increasing energy consumption, it can get a higher reward; conversely, if the energy consumption is too high and the filtration efficiency is not significantly improved, it will get a lower reward. In this way, the purifier is prompted to find a combination of working parameters that can both improve filtration efficiency and reduce energy consumption.

[0134] Using GNN for graph structure modeling, specifically using GNN to model the topological relationship between system parameters and pollution feature knowledge base, can better capture the complex associations between data. Using working parameters and pollution features as nodes of the graph and their mutual relationships as edges, GNN can use the message passing mechanism to allow nodes to aggregate information from neighboring nodes, learn richer feature representations, and provide more accurate basis for subsequent decision-making.

[0135] S32. Construct a GNN-based policy network Actor and a value function network Critic to dynamically generate parameter optimization strategies;

[0136] The policy network Actor outputs an action probability distribution based on the current state information, that is, the probability of taking each action in the current state. The Actor network built on GNN can make full use of the information contained in the graph structure to effectively extract and analyze the complex state, thereby generating action strategies more reasonably. When faced with different pollution space distributions and working parameter combinations, the Actor network can quickly give appropriate action suggestions, such as which working parameters to adjust and the extent of the adjustment.

[0137] The value function network Critic is used to evaluate the value of the current state, that is, the expected long-term cumulative reward after executing a series of actions starting from the current state. It estimates the value of the state through learning and provides feedback to the policy network Actor. The Critic network can help the Actor network determine whether an action is in the direction of maximizing the long-term cumulative reward, thereby guiding the Actor network to continuously optimize the strategy. When the Actor network takes an action, the Critic network will evaluate the impact of this action on the state value and feedback to the Actor network, allowing the Actor network to adjust subsequent action strategies.

[0138] S33. Use the policy gradient algorithm to update the policy network and maximize the long-term cumulative rewards. The goal of the policy gradient algorithm is to maximize the long-term cumulative rewards. It adjusts the parameters of the policy network by calculating the gradient of the policy network parameters so that the policy network can obtain higher rewards when performing actions. Specifically, the policy gradient algorithm calculates the gradient of the policy network parameters based on the rewards obtained after the current policy executes the action, and then updates the parameters in the direction of the gradient ascent. In each iteration, the action is selected and executed according to the current policy network Actor, and the feedback from the environment is observed to obtain rewards and new states. Using this information, the gradient of the policy network is calculated, and the parameters of the policy network are updated using an optimizer (such as the Adam optimizer). As the iterations proceed, the policy network will gradually converge to a strategy that can maximize the long-term cumulative rewards, that is, find a working parameter combination strategy library that maximizes filtering efficiency.

[0139] In this embodiment, the step S4 uses multi-physics field coupling simulation software to simulate and analyze the plasma air purification system, associates the simulation results with the reinforcement learning strategy library, and generates a parameter optimization knowledge graph, specifically:

[0140] S41. Establish an electric field-flow field-concentration field coupling simulation model of the plasma purification system, including electric field control equations, flow field control equations, and pollutant control equations, where:

[0141] The governing equation for the electric field is: φ is the electric potential, ε is the dielectric constant, and ρ is the space charge density;

[0142] The governing equations of the flow field are:

[0143]

[0144] u describes the velocity of a point in the fluid at a certain moment, including the magnitude and direction of the velocity; t indicates the order of occurrence and duration of the physical process; p is the force acting vertically on the unit area of ​​the fluid; u is the dynamic viscosity of the fluid; F e is the electric force;

[0145] Pollutant transport equation:

[0146]

[0147] C refers to the concentration of pollutants in space, that is, the content of pollutants per unit volume; D is the diffusion coefficient. The larger the diffusion coefficient, the faster the pollutants diffuse; k represents the speed of the reaction between plasma and pollutants;

[0148] Define the simulation input parameter x sim=[V, u, d, T, ε], V, u, d, T, ε represent voltage, flow rate, electrode spacing, temperature, medium properties, and output performance index y sim =[η, P, ΔC], η, P, ΔC represent filtration efficiency, energy consumption, and concentration gradient, respectively;

[0149] S42, associate the simulation data with the strategy library, map the parameter combination in the reinforcement learning strategy library to the simulation input, calculate the difference between the simulation result and the expected performance of the strategy library, and if the difference is less than the threshold, mark the parameter combination as a credible strategy, otherwise trigger simulation calibration. The specific operation process is as follows:

[0150] Take out the parameter combination from the reinforcement learning strategy library, and input these parameters one by one into the plasma purification system model established by the multi-physics field coupling simulation software. Set the parameters such as discharge voltage, wind speed, and electrode spacing accurately to the corresponding positions of the simulation model to ensure that the simulation environment is consistent with the parameter settings in the strategy library. After setting the parameters, run the simulation software to simulate the working process of the plasma air purifier under this parameter combination. Obtain key performance indicators in the simulation results, such as filtration efficiency, energy consumption, concentration gradient and other data. Compare the performance indicators obtained by simulation with the expected performance in the strategy library and calculate the difference between the two. The degree of difference can be measured by methods such as mean square error and absolute error. Set a reasonable threshold. If the calculated difference is less than the threshold, the parameter combination is determined to be a credible strategy, marked and stored for subsequent use; if the difference is greater than the threshold, the simulation calibration process is triggered.

[0151] S43, build a heterogeneous knowledge graph, including three types of nodes and edges, where the nodes include parameter nodes (x sim =[V,u,d,T,ε]), performance node (y sim =[η, P, ΔC]), strategy node (π 1 , π 2 ,…π n ), and the edge types include causal relationship, strategy association, and simulation verification. The knowledge graph here integrates rich information and provides an intuitive and comprehensive basis for the control decision of the plasma air purifier. By querying the graph, we can quickly understand the impact of different parameters on performance and which strategies may achieve the expected performance goals, thereby guiding the actual control operation.

[0152] The parameter nodes cover various key parameters of the plasma air purifier operation, such as the discharge voltage of the needle electrode array, the electrostatic field strength between the metal foam and the grounded mesh electrode, the wind speed in the ventilation duct, the distance between the needle electrode array and the metal foam, and the PPI of the metal foam. These parameters are the basic variables for the operation of the purifier and directly affect its performance. Each parameter node carries detailed attribute information, including the value of the parameter, measurement time, measurement location, etc., in order to accurately describe the state of the parameter.

[0153] Performance nodes are mainly used to characterize the working performance of the purifier, such as filtration efficiency, energy consumption, concentration gradient, etc. These performance indicators reflect the actual performance of the purifier under different working conditions. The filtration efficiency node records the removal ratio of different pollutants, and the energy consumption node reflects the energy consumption during operation. The performance node is associated with the parameter node to reveal the impact of parameter changes on the performance of the purifier.

[0154] The strategy node stores the optimization working parameter combination strategy generated by algorithms such as reinforcement learning. Each strategy node corresponds to a specific set of parameter settings, which are formulated to achieve goals such as maximizing filtering efficiency or minimizing energy consumption. The strategy node is closely connected to the parameter node, which clarifies the specific parameter values ​​corresponding to the strategy, and is associated with the performance node to evaluate the effect of the strategy after implementation.

[0155] Causal edges are used to connect parameter nodes and performance nodes, reflecting the direct impact of operating parameters on purifier performance. Changes in discharge voltage will affect the electrostatic field strength, which in turn affects the ionization and adsorption of pollutants, and ultimately affects the filtration efficiency. This relationship can be represented by causal edges. The weight of the causal edge can be set according to the degree of influence to quantify the strength of this relationship.

[0156] The policy association edge establishes the connection between the policy node and the parameter node, indicating the specific parameter combination corresponding to a specific policy. A policy association edge starts from a policy node and connects to the corresponding parameter node, clearly showing the values ​​of each parameter involved in the policy. This helps to quickly find and apply different policies.

[0157] The simulation verification edge is used to connect the strategy node and the performance node. After the parameter combination in the strategy library is input into the simulation model for simulation operation and the simulation results are obtained, the simulation verification edge is used to represent the verification relationship between the strategy and the corresponding performance. If the simulation results show that a certain strategy can effectively improve the filtering efficiency, then the simulation verification edge from the strategy node to the filtering efficiency performance node represents the effectiveness of the strategy at the simulation level.

[0158] S44. Generate node embedding using graph attention network GAT

[0159]

[0160] represents the embedding representation of node i in the l+1th layer of the graph attention network; σ is the activation function; It means to sum the neighbor node set N(i) of node i, where N(i) includes all nodes directly connected to node i;

[0161] α ij is the attention coefficient, which is used to measure the degree of association between node i and its neighbor node j; W (l) is the weight matrix of the lth layer; represents the embedding representation of neighbor node j at layer l.

[0162] The graph attention network generates a new node embedding representation by aggregating the information of the node itself and its neighboring nodes through the attention mechanism. In the formula, the node embedding representation of the new layer is It is obtained by weighted summing of the node i’s neighbor node set N(i) and processing it with the activation function σ. This method can focus on aggregating information according to the degree of association between nodes, which is more flexible and effective than traditional graph neural networks.

[0163] In this embodiment, step S5 generates a plasma air purifier control instruction according to the knowledge graph to control the working state of the plasma air purifier, specifically:

[0164] S51. Obtain real-time data about the environment and equipment through sensors and map them into state vectors that can be recognized by the knowledge graph. Use a variety of sensors, such as laser dust sensors, temperature and humidity sensors, VOC sensors, micro laser radar (LiDAR), etc., to distribute them in a grid pattern inside and around the purifier. These sensors collect information including pollutant concentration (PM2.5, PM10, VOC, etc.), temperature and humidity, CO 2 , as well as multi-dimensional data such as equipment operating parameters (such as discharge voltage, wind speed, and electrode spacing). The collected data needs to be preprocessed to convert it into a form suitable for subsequent processing. The preprocessed data is organized and encoded according to the rules defined by the knowledge graph and mapped into a state vector that can be recognized by the knowledge graph. The nodes in the knowledge graph include parameter nodes (such as voltage, flow rate, electrode spacing, etc.), performance nodes (such as filtration efficiency, energy consumption, etc.) and strategy nodes. Each sensor data corresponds to the corresponding node attribute to form a vector representation describing the current environment and equipment status. For example, the PM2.5 concentration, temperature, discharge voltage and other data collected at a certain moment are respectively corresponded to the attributes of the corresponding pollution spatial distribution feature nodes and environmental parameter nodes in the knowledge graph, thereby forming a state vector that can reflect the current system state.

[0165] S52. Based on the state vector, the optimal control strategy matching the current state is retrieved in the knowledge graph by using the cosine similarity matching method. The specific process is as follows:

[0166] Constructing a knowledge graph: The knowledge graph contains parameter nodes (such as voltage, flow rate, electrode spacing, etc.), performance nodes (such as filtration efficiency, energy consumption, etc.), and strategy nodes. These nodes are interconnected through causal edges, strategy-related edges, and simulation verification edges to form an organic whole that comprehensively stores and represents various types of information about plasma air purifiers. For example, causal edges reflect the impact of working parameters on purifier performance, strategy-related edges show the parameter combinations corresponding to specific strategies, and simulation verification edges verify the relationship between strategy and performance.

[0167] Obtaining state vector: Real-time data of the environment and equipment, such as pollutant concentration, temperature and humidity, and equipment operating parameters, are obtained through sensors. After preprocessing such as data cleaning and normalization, they are mapped into state vectors that can be recognized by the knowledge graph. This vector fully describes the current operating status of the plasma air purifier and is the basis for subsequent matching.

[0168] Cosine similarity matching principle: Cosine similarity measures the similarity of two vectors by calculating the cosine value of the angle between them. The value range is between [-1, 1]. The closer the value is to 1, the more similar the two vectors are. In this system, the cosine similarity is calculated between the state vector and the state vector associated with each policy node in the knowledge graph. For example, assuming there are policies A and B, corresponding to state vectors S_A and S_B respectively, and the current state vector is S_C, the cosine similarities Sim(S_C, S_A) and Sim(S_C, S_B) between S_C and S_A, and S_C and S_B are calculated respectively.

[0169] Retrieve the optimal control strategy: Sort the calculated similarities and select the strategy with the highest similarity as the optimal control strategy that matches the current state. This strategy is based on a large amount of empirical data and optimization results in the knowledge graph, and can provide the best control parameter combination for the current operating state of the plasma air purifier, thereby achieving efficient purification effects and lower energy consumption. If Sim(S_C, S_A)>Sim(S_C, S_B), and Sim(S_C, S_A) is the largest among all calculated similarities, then strategy A is the optimal control strategy under the current state.

[0170] S53, converting the strategy parameters into physical control signals executable by the device, specifically using a PID control algorithm, and controlling the voltage of the plasma generator by adjusting the pulse width modulation PWM signal. The calculation formula is:

[0171] PWM v =Kp ·(V opt -V real )+K i ·∫(V opt -V real )dt

[0172] Among them, PWM v K is a pulse width modulation (PWM) signal used to control the voltage of the plasma generator; p is the proportionality coefficient, K i is the integral coefficient; V opt is the optimal voltage value, V real is the actual measured voltage value of the plasma generator. When the plasma air purifier is running, different air quality conditions require the plasma generator to output a suitable voltage to generate plasma of appropriate intensity to purify the air. By adjusting the PWM signal and then controlling the voltage through the PID control algorithm, the purifier can maintain efficient purification performance when facing different pollution conditions.

[0173] S54. Send instructions to the plasma generator execution unit through the industrial bus protocol. In a plasma generator, a cathode and an anode are usually provided. When a sufficiently high voltage is applied between the two poles, a small amount of free electrons in the gas will be accelerated under the action of the electric field and collide with gas molecules. These collisions will cause the electrons in the gas molecules to break away from the atoms, generating more free electrons and positive ions, forming a gas discharge phenomenon and generating an arc. Gas ionization methods include thermal ionization, electric field ionization, and photoionization. In a plasma generator, electric field ionization is mainly used, that is, a strong electric field is used to make gas particles obtain enough energy, causing electrons to break away from atoms or molecules to form plasma. As the ionization process continues, the gas will contain a large number of particles such as electrons, ions, atoms and molecules, and the material state composed of these particles is plasma. In order to obtain more energetic and active plasma, the plasma arc will be subjected to three compression effects during the formation process, namely mechanical compression effect (compression by ejecting the arc through a channel with a smaller aperture), thermal compression effect (using water cooling to cool the gas near the anode hole wall, forcing the arc to compress toward the center) and self-magnetic compression effect (the magnetic field generated by the arc's own current causes the arc to be compressed toward the center). These compression effects make the energy of the plasma arc more concentrated, the temperature higher, and the flow rate faster. The control instructions of the plasma generator execution unit usually include the following categories:

[0174] Power control instructions: used to control the power on / off, voltage, current and other parameters of the plasma generator to ensure stable operation and generate plasma of appropriate intensity. For example, the energy level of the plasma can be changed by adjusting the voltage instruction to meet different application requirements.

[0175] Working mode command: Set the working mode of the plasma generator, such as continuous working mode, pulse working mode, etc. Different working modes are suitable for different scenarios. For example, in air purification, the continuous mode can maintain continuous purification, while the pulse mode can save energy to a certain extent.

[0176] Gas control instructions: When a specific gas is required to participate in the plasma generation process, the execution unit will receive instructions such as gas type selection and gas flow adjustment. For example, in some industrial processes, it may be necessary to accurately control the input amount of gases such as oxygen and argon.

[0177] Status monitoring and feedback instructions: The execution unit not only receives control instructions, but also feeds back its own working status information to the control system, such as temperature, pressure, fault signals, etc., so that the control system can adjust instructions or issue alarms in time.

[0178] S55. Combine sensor feedback to correct control parameters in real time to form a closed-loop optimization. When the sensor feedback data deviates from the preset ideal operating state, the control system will dynamically adjust the control parameters using the PID control algorithm according to the size and direction of the deviation. For example, if an increase in PM2.5 concentration is detected in the air, the system will automatically increase the voltage of the plasma generator to enhance the purification capacity.

[0179] In this embodiment, the step S6 iteratively detects the spatial distribution of pollution, and dynamically corrects the reinforcement learning strategy through difference analysis between twin data and real data, specifically:

[0180] S61. Use sensors to continuously collect real data on the spatial distribution of pollution in and around the plasma air purifier. In order to comprehensively and accurately collect data, multiple types of sensors need to be selected, such as laser dust sensors to monitor PM2.5 and PM10 concentrations, VOC sensors to detect volatile organic compounds, and temperature and humidity sensors to record ambient temperature and humidity. They can reflect the impact of pollution characteristics and environmental factors on purification. The layout of sensors is also very critical. They can be installed at the air inlet, ionization zone, filter module and other locations inside the purifier to monitor the pollution status at different purification stages; they can be evenly distributed in a grid around the periphery to grasp the spatial distribution of pollution.

[0181] S62. Based on the established pollution feature knowledge base, system model and previously accumulated data, digital twin technology is used to generate twin data corresponding to the real scene. The following is the detailed implementation process of using digital twin technology to generate twin data:

[0182] 1. Build a digital twin basic framework

[0183] The spatial distribution characteristics of pollution and working parameter characteristics in the pollution characteristic knowledge base are integrated with the system model (such as the purifier model with multi-physical field coupling), and the historically accumulated sensor data, control strategy execution data, etc. are connected to form the data and model basis of the digital twin.

[0184] Clarify the mapping rules between the real scene (plasma air purifier and surrounding polluted space) and the virtual model. For example, correspond physical quantities such as pollutant concentration and equipment operating parameters in the real environment to the node attributes or parameter variables of the virtual model.

[0185] 2. Data access and real-time drive

[0186] Through sensors, the pollution concentration inside the purifier (such as the ionization zone and filter module) and the surrounding environment, as well as the equipment operating parameters (voltage, wind speed, etc.) are continuously acquired and transmitted to the digital twin system in real time.

[0187] The accumulated historical data (such as pollution control effect data under different working conditions and historical equipment failure data) are imported into the virtual model to train the model's simulation capabilities for complex scenarios, so that the twin data can not only reflect the real-time status but also reproduce historical laws.

[0188] 3. Virtual model simulation and twin data generation

[0189] Based on the system model, the rules in the pollution feature knowledge base (such as pollution diffusion model and purification efficiency formula) are used to calculate the real-time access data. For example, based on the current wind speed, pollutant concentration and purifier working parameters, the diffusion path and purification process of pollution in space are simulated.

[0190] In the virtual space, twin data corresponding to the real scene are output, including virtual pollution concentration distribution cloud maps, virtual equipment operation parameter curves, virtual predicted values ​​of purification efficiency, etc., to fully reproduce the status and changing trends of the real scene.

[0191] 4. Twin data verification and iterative optimization

[0192] Compare the real sensor data with the twin data, and evaluate the accuracy of the twin data through error analysis (such as mean square error, absolute error). If the error exceeds the threshold, optimize the system model parameters or modify the pollution feature knowledge base rules.

[0193] As real-scene data continues to accumulate, the digital twin model is regularly updated and the simulation algorithm is optimized to ensure that the twin data always accurately maps the real scene, providing reliable data support for subsequent control strategy optimization, fault prediction and other applications.

[0194] S63. Normalize the collected real data and the generated twin data, and use the mean square error Euclidean distance indicator to calculate the difference between the twin data and the real data. In the application of plasma air purifier, if there are real PM2.5 concentration data and corresponding twin data collected at 10 time points, the difference between the real value and the twin value at each time point is squared and summed, and then divided by 10 to obtain the mean square error of this set of data. The mean square error can intuitively reflect the degree of difference in the overall data. The smaller the value, the closer the twin data is to the real data. Euclidean distance measures the straight-line distance between two points in multidimensional space. When calculating the difference between twin data and real data, different types of data (such as PM2.5 concentration, temperature, humidity, etc.) at each time point are regarded as different dimensions of a multidimensional vector, and the difference between the two multidimensional vectors is measured by calculating the Euclidean distance. Euclidean distance not only takes into account the difference in data, but also takes into account the comprehensive influence between different dimensions, which can more comprehensively reflect the difference between data.

[0195] S64. Adjust the reward function in reinforcement learning according to the dimension with significant difference. Assuming that the original reward function is R(s,a), the adjustment factor based on the difference is α, and the adjusted reward function is R`(s,a), then:

[0196] R`(s,a)=αR(s,a), where a is the action and s is the state;

[0197] S65. Use a reinforcement learning algorithm such as a policy gradient algorithm and use the adjusted reward function to update the policy network. The update formula of the policy network parameter θ is:

[0198]

[0199] Among them, θ t is the parameter of the current policy network, η is the learning rate, π θ (a|s) is the probability of taking action a in state s;

[0200] S66. Integrate the newly collected real data and the calculated difference information into the pollution feature knowledge base, and update and optimize the pollution data distribution pattern and working parameter characteristics. The newly collected real data covers various information about the inside and surrounding environment of the air purifier, such as the pollutant concentrations (PM2.5, PM10, VOC, etc.) at different locations and the real-time working parameters of the equipment (discharge voltage, wind speed, electrode spacing, etc.). Before being integrated into the knowledge base, these data need to be strictly cleaned and verified to remove outliers and erroneous data to ensure the accuracy and reliability of the data. The calculated difference information refers to indicators such as the mean square error and Euclidean distance between the twin data and the real data. These difference information reflects the degree of deviation between the digital twin model and the actual situation. Conduct an in-depth analysis of the difference information to find out the reasons for the difference, which may be inaccurate model parameters, incomplete consideration of environmental factors, or data collection errors. Integrate the verified new real data and the analyzed difference information into the pollution feature knowledge base. For new data, store it in the corresponding node or relationship according to the structure and format of the knowledge base. For difference information, it can be associated with related data as additional attributes or metadata.

[0201] S67. According to the results of the strategy correction, fine-tune the parameters of the graph neural network model, reinforcement learning model, and simulation model. For example, if the strategy correction finds that the current model has insufficient processing capabilities for complex topological structures, the structural parameters such as the number of layers of the graph neural network and the number of neurons in each layer can be fine-tuned. Increase the number of hidden layers or the number of neurons to improve the expressive power of the model so that it can more accurately capture the high-order relationships between nodes in the graph, thereby providing a more reliable feature basis for strategy generation. For example, in the simulation model of a plasma purification system with multi-physical field coupling, adjust parameters such as electric field strength and ion mobility to make it more consistent with actual physical phenomena, thereby improving the simulation accuracy of the simulation model for the actual system.

[0202] like Figure 2 As shown, the present invention also proposes a plasma air purifier control system based on artificial intelligence, comprising:

[0203] The pollution feature knowledge base construction module 101 collects the working parameters of the plasma air purifier system, extracts the pollution spatial distribution characteristics in the plasma air purifier through the improved PointNet++ network, and constructs the pollution feature knowledge base;

[0204] A graph neural network model building module 102 is used to connect the pollution feature knowledge base to the graph neural network model, use the graph neural network model to model and train the collected working parameters and the system's filtration efficiency for aerosols, and obtain a graph neural network model that can accurately predict filtration efficiency;

[0205] A strategy library construction module 103 is used to optimize the working parameters of the system by using a reinforcement learning algorithm combined with a trained graph neural network model to form a strategy library of working parameter combinations that maximize filtering efficiency;

[0206] The knowledge graph construction module 104 is used to use multi-physics field coupling simulation software to perform simulation analysis on the plasma air purification system, associate the simulation results with the reinforcement learning strategy library, and generate a parameter optimization knowledge graph;

[0207] A control instruction generation module 105 is used to generate a plasma air purifier control instruction according to the knowledge graph to control the working state of the plasma air purifier;

[0208] The strategy optimization module 106 is used to iteratively detect the spatial distribution of pollution and dynamically correct the reinforcement learning strategy through difference analysis between twin data and real data.

[0209] The technical principle of the system is:

[0210] 1. Pollution feature perception and modeling

[0211] With the help of a variety of high-precision sensors, such as laser dust sensors, VOC sensors, temperature and humidity sensors, etc., a dense sensor network is built to comprehensively collect data from the inside and surrounding environment of the plasma air purifier. These sensors are carefully deployed at the air inlet, ionization zone, filter module and different spatial locations around the purifier to obtain multi-dimensional data including pollutant concentration, temperature and humidity, voltage, wind speed, etc. The collected data will first be cleaned to remove outliers caused by sensor failure or environmental interference, and then use data standardization methods such as Min-Max normalization to unify data of different dimensions into a specific range, laying the foundation for subsequent analysis.

[0212] The collected pollution spatial distribution data is processed using the improved PointNet++ network. First, the sensor data is converted into point cloud data containing location and pollution information. Key points are evenly selected in the point cloud space through farthest point sampling, and then the nearest K points are found for each key point to form a local neighborhood. The PointNet module is responsible for extracting these local neighborhood features, calculating the eigenvalues ​​of the current layer points through inverse distance weighted interpolation, and concatenating them with the original features. After processing by a multi-layer perceptron (MLP), a more representative new feature matrix is ​​obtained. After multi-layer processing, the network outputs the global feature vector of each point. Finally, through average pooling or maximum pooling operations, a vector that can reflect the global characteristics of the pollution space is obtained, thereby accurately capturing the spatial distribution pattern of pollution.

[0213] The principal component analysis (PCA) is used to reduce the dimension of the pre-processed working parameters, while retaining the main information and reducing data redundancy. The clustering algorithm K-Means is used to continuously calculate the distance between the sample and the cluster center and adjust the center to achieve cluster analysis of the pollution data, thereby discovering the distribution pattern of the pollution data and distinguishing areas with different pollution levels. The graph database Neo4j is used to store the extracted features of the working parameters and the pollution space distribution features. Nodes representing various entities are created in the database, such as working parameter nodes and pollution space distribution feature nodes. Causal edges, spatiotemporal correlation edges, and equipment correlation edges are created based on the relationship between them. Data is collected and updated periodically to build a comprehensive and dynamic pollution feature knowledge base.

[0214] 2. Graph Neural Network Modeling and Prediction

[0215] Extract data from the pollution feature knowledge base, normalize the numerical data to make different features comparable; use unique hot encoding for categorical data to convert it into a computer-recognizable vector form. Use pollution features and working parameters as nodes, and construct an adjacency matrix based on the relationship between them. If there is an edge between node i and node j, the corresponding element in the adjacency matrix is ​​1, otherwise it is 0. At the same time, combine the preprocessed data into a feature matrix to complete the construction of the graph data structure.

[0216] The graph convolutional network (GCN) is selected as the modeling tool. The GCN layer uses a specific calculation formula to perform convolution operations on node features using the adjacency matrix and degree matrix, so that the model can learn the local and global relationships between nodes. By stacking multiple GCN layers, the model can capture more complex graph structure features. After the last layer of GCN, a fully connected layer is added as the output layer to map the node features to the predicted filtering efficiency. The mean square error (MSE) loss function is used to measure the difference between the filtering efficiency predicted by the model and the actual filtering efficiency. The model prediction value is calculated through forward propagation, and the loss value is calculated according to the loss function. Then, the optimizer (such as the Adam optimizer) is used to update the model parameters through back propagation. The training is continuously iterated until the loss value converges or the preset number of training rounds is reached, thereby obtaining a graph neural network model that can accurately predict the filtering efficiency.

[0217] (III) Reinforcement Learning Strategy Optimization

[0218] The operating environment of the plasma air purifier is abstracted as a Markov decision process, and the state space, action space and reward function are clearly defined. The state space contains key information such as working parameters (such as discharge voltage, electrostatic field strength, wind speed, etc.) and pollution spatial distribution characteristics, which comprehensively describes the current operating state of the purifier; the action space is the adjustment operation of the working parameters, such as changing the discharge voltage and adjusting the wind speed; the reward function uses the improvement of filtration efficiency as a positive reward and the increase of energy consumption as a negative reward, guiding the purifier to find a combination of working parameters that can both improve filtration efficiency and reduce energy consumption.

[0219] Based on the graph neural network (GNN), the policy network Actor and the value function network Critic are constructed. The policy network Actor outputs the action probability distribution based on the current state information and determines the probability of taking each action in the current state. With the help of GNN's ability to process graph structured data, the Actor network can effectively extract and analyze complex states and reasonably generate action strategies. The value function network Critic is used to evaluate the value of the current state, that is, the expected long-term cumulative reward after performing a series of actions starting from the current state. The Critic network continuously estimates the state value through learning, provides feedback to the Actor network, helps it determine whether the action is in the direction of maximizing the long-term cumulative reward, and thus guides the Actor network to optimize the strategy.

[0220] The policy network is updated using a policy gradient algorithm, the goal of which is to maximize long-term cumulative rewards. In each iteration, the action is selected and executed based on the current policy network Actor, and the reward and new state are obtained by observing the environmental feedback. This information is used to calculate the gradient of the policy network, and the optimizer (such as the Adam optimizer) is used to update the parameters of the policy network in the direction of the gradient rise. As the iteration proceeds, the policy network gradually converges to a strategy that can maximize the long-term cumulative rewards, forming a working parameter combination strategy library that maximizes filtering efficiency.

[0221] (IV) Multi-physics field coupling simulation and knowledge graph

[0222] Using multi-physics coupling simulation software, an electric field-flow field-concentration field coupling simulation model of the plasma purification system is established. The model contains electric field control equations, flow field control equations, and pollutant control equations, which respectively describe the changing laws of physical quantities such as electric potential, electric field force, fluid velocity, and pollutant concentration. Define simulation input parameters such as voltage, flow rate, electrode spacing, temperature, medium properties, etc., as well as output performance indicators such as filtration efficiency, energy consumption, and concentration gradient. By solving these equations and setting parameters, the purification process of the plasma air purifier under different working conditions is simulated.

[0223] Map the parameter combinations in the reinforcement learning strategy library to the simulation input, simulate the working process of the plasma air purifier in the simulation software, and obtain the key performance indicators in the simulation results. Compare the simulation results with the expected performance of the strategy library, and use methods such as mean square error and absolute error to measure the degree of difference. If the difference is less than the preset threshold, mark the parameter combination as a credible strategy; if the difference is greater than the threshold, trigger the simulation calibration, adjust the simulation model parameters, and re-simulate and verify to ensure the reliability of the parameter combination in the strategy library.

[0224] Construct a heterogeneous knowledge graph containing parameter nodes, performance nodes, and strategy nodes. Parameter nodes cover various operating parameters of plasma air purifiers, and each node carries detailed attributes such as parameter value, measurement time, and measurement location; performance nodes characterize the working effectiveness of the purifier, such as filtration efficiency, energy consumption, etc.; strategy nodes store the optimization working parameter combination strategy generated by reinforcement learning. The edges in the knowledge graph include causal edges, which reflect the direct impact of working parameters on purifier performance; strategy association edges, which show the specific parameter combination corresponding to a specific strategy; and simulation verification edges, which are used to verify the relationship between the strategy and the corresponding performance. Through the knowledge graph, rich information is integrated to provide an intuitive and comprehensive basis for control decisions.

[0225] 5. Digital Twins and Closed-Loop Correction

[0226] Based on the established pollution feature knowledge base, system model and historical accumulated data, digital twin technology is used to build a digital twin model. Physical quantities such as pollutant concentration and equipment operating parameters in the real scene are mapped to the node attributes or parameter variables of the virtual model to establish a mapping relationship between the real scene and the virtual model. Real-time data of the purifier's internal and surrounding environment is continuously obtained through sensors and transmitted to the digital twin system. Historical data is imported to train the virtual model so that it can simulate complex scenes and generate twin data corresponding to the real scene, such as virtual pollution concentration distribution cloud maps, equipment virtual operation parameter curves, etc.

[0227] The collected real data and the generated twin data are normalized, and the mean square error Euclidean distance indicator is used to calculate the difference between the two. The reward function in reinforcement learning is adjusted according to the difference significance dimension to adapt to environmental changes and model deviations. Reinforcement learning algorithms such as the policy gradient algorithm are used to update the policy network using the adjusted reward function. The newly collected real data and the calculated difference information are integrated into the pollution feature knowledge base to update and optimize the pollution data distribution pattern and working parameter characteristics. According to the policy correction results, the parameters of the graph neural network model, reinforcement learning model, and simulation model are fine-tuned to achieve dynamic optimization of the control strategy and continuous evolution of the system.

[0228] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A plasma air purifier control method based on artificial intelligence, characterized in that: The following steps are involved: The working parameters of the plasma air purifier system were collected, and the pollution spatial distribution characteristics in the plasma air purifier were extracted through the improved PointNet++ network to build a pollution feature knowledge base. The pollution feature knowledge base is connected to the graph neural network model, and the graph neural network model is used to model and train the collected working parameters and the system's filtration efficiency for aerosols to obtain a graph neural network model that can accurately predict filtration efficiency. The reinforcement learning algorithm is combined with the trained graph neural network model to optimize the system's working parameters and form a strategy library for the working parameter combination that maximizes filtering efficiency. Use multi-physics field coupling simulation software to simulate and analyze the plasma air purification system, associate the simulation results with the reinforcement learning strategy library, and generate a parameter optimization knowledge graph; Generate plasma air purifier control instructions based on the knowledge graph to control the working state of the plasma air purifier; Iteratively detect the spatial distribution of pollution, and dynamically correct the reinforcement learning strategy through difference analysis between twin data and real data.

2. The artificial intelligence-based plasma air purifier control method according to claim 1, characterized in that: The improved PointNet++ network is used to extract the pollution spatial distribution characteristics in the plasma air purifier, specifically: Use sensors to collect spatial distribution data of pollution in and around the plasma air purifier, convert it into point cloud data containing location and pollution information, and normalize each feature dimension of the data; The improved PointNet++ network structure is used for hierarchical sampling, feature extraction and feature propagation. The sampling layer uses the farthest point sampling to select key points from the current layer point cloud, and the grouping layer finds the nearest K points for each key point in the current layer point cloud to form a local neighborhood. Then the PointNet module is used to extract local neighborhood features, and the feature values ​​are calculated for the current layer points through inverse distance weighted interpolation. The interpolated features are concatenated with the original features of the current layer, and the new feature matrix is ​​obtained through MLP processing. After multi-layer processing, the network outputs the global feature vector of each point, and then obtains the global features of the contaminated space through average pooling or maximum pooling.

3. The artificial intelligence-based plasma air purifier control method according to claim 1, characterized in that: The construction of the pollution feature knowledge base specifically includes: Preprocess the collected working parameters for data cleaning and data standardization; The principal component analysis was used to reduce the dimension and extract the principal component. The clustering algorithm K-Means was used to calculate the distance between the sample and the cluster center and continuously adjust the center to achieve clustering, discover the distribution pattern of pollution data, and distinguish areas with different pollution levels. The graph database Neo4j is used to store the extracted features of the working parameters and the pollution space distribution features. Nodes representing entities and edges representing entity relationships are created in the graph database, and data are collected and updated periodically to complete data construction.

4. The artificial intelligence-based plasma air purifier control method according to claim 1, characterized in that: The pollution feature knowledge base is connected to the graph neural network model, and the graph neural network model is used to model and train the collected working parameters and the system's filtration efficiency for aerosols, specifically: Extract data from the pollution feature knowledge base, normalize numerical data and perform one-hot encoding on categorical data during preprocessing; Taking pollution characteristics and working parameters as nodes, edges are built according to the association relationship, and the adjacency matrix A is constructed accordingly. If there is an edge between node i and node j, then A ij =1, otherwise A ij =0; at the same time, the preprocessed data is combined into a feature matrix X to complete the construction of the graph data structure; Select the graph neural network model GCN, the calculation formula of the GCN layer is: Among them, H (l) is the node feature matrix of the lth layer, W (l) is the weight matrix of the lth layer, σ is the activation function, is the adjacency matrix with self-connection added, I is the identity matrix, yes The degree matrix, whose elements By stacking multiple GCN layers, the model can learn more complex graph structure features. After the last layer of GCN, a fully connected layer is added as the output layer to map the node features to the predicted filtering efficiency. Assume that the node feature matrix after multi-layer GCN processing is H (L) , the calculation formula of the output layer is: y=FC(H (L) ) Where y is the predicted filtering efficiency, FC represents the fully connected layer operation; The mean square error (MSE) loss function is used to measure the difference between the filtration efficiency predicted by the model and the actual filtration efficiency. The loss function formula is: Where n is the number of samples, y i is the actual filtration efficiency, is the filtration efficiency predicted by the model. Input the constructed graph data into the model, calculate the predicted value of the model through forward propagation, then calculate the loss value according to the loss function, use the optimizer to update the model parameters through back propagation, and continuously iterate the training until the loss value converges or reaches the preset number of training rounds.

5. The artificial intelligence-based plasma air purifier control method according to claim 1, characterized in that: The reinforcement learning algorithm is combined with the trained graph neural network model to optimize the working parameters of the system to form a strategy library of working parameter combinations that maximize filtering efficiency, specifically: The operating environment of the plasma air purifier is modeled as a Markov decision process, the state space, action space and reward function are defined, and GNN is used to perform graph structure modeling on the topological relationship between system parameters and pollution feature knowledge base; Construct a GNN-based policy network Actor and a value function network Critic to dynamically generate parameter optimization strategies; The policy gradient algorithm is used to update the policy network to maximize the long-term cumulative reward.

6. The artificial intelligence-based plasma air purifier control method according to claim 1, characterized in that: The multi-physics field coupling simulation software is used to simulate and analyze the plasma air purification system, and the simulation results are associated with the reinforcement learning strategy library to generate a parameter optimization knowledge graph, specifically: (1) Establish an electric field-flow field-concentration field coupling simulation model of the plasma purification system, including the electric field control equation, flow field control equation, and pollutant control equation, where: The governing equation for the electric field is: ▽·(ε▽φ)=-ρ, where φ is the electric potential, ε is the dielectric constant, and ρ is the space charge density; The governing equations of the flow field are: u describes the velocity of a point in the fluid at a certain moment, including the magnitude and direction of the velocity; t indicates the order of occurrence and duration of the physical process; p is the force acting vertically on the unit area of ​​the fluid; u is the dynamic viscosity of the fluid; F e is the electric force; Pollutant transport equation: C refers to the concentration of pollutants in space, that is, the content of pollutants per unit volume; D is the diffusion coefficient. The larger the diffusion coefficient, the faster the pollutants diffuse; k represents the speed of the reaction between plasma and pollutants; Define the simulation input parameter x sim =[V, u, d, T, ε], V, u, d, T, ε represent voltage, flow rate, electrode spacing, temperature, medium properties, and output performance index y sim =[η, P, ΔC], η, P, ΔC represent filtration efficiency, energy consumption, and concentration gradient, respectively; (2) Associating simulation data with the policy library, mapping the parameter combination in the reinforcement learning policy library to the simulation input, and calculating the difference between the simulation result and the expected performance of the policy library. If the difference is less than a threshold, the parameter combination is marked as a credible policy, otherwise simulation calibration is triggered; (3) Construct a heterogeneous knowledge graph, which contains three types of nodes and edges, where the nodes include parameter nodes (x sim =[V,u,d,T,ε]), performance node (y sim =[η,P,ΔC]), strategy node (π1, π2,…π n ), edge types include causal relationship, policy association, and simulation verification; (4) Generate node embedding using graph attention network GAT represents the embedding representation of node i in the l+1th layer of the graph attention network; σ is the activation function; It means to sum the neighbor node set N(i) of node i, where N(i) includes all nodes directly connected to node i; α ij is the attention coefficient, which is used to measure the degree of association between node i and its neighbor node j; W (l) is the weight matrix of the lth layer; represents the embedding representation of neighbor node j at layer l.

7. The artificial intelligence-based plasma air purifier control method according to claim 1, characterized in that: The plasma air purifier control instructions are generated according to the knowledge graph to control the working state of the plasma air purifier, specifically: Acquire real-time data of the environment and equipment through sensors and map them into state vectors that can be recognized by the knowledge graph; Based on the state vector, the optimal control strategy matching the current state is retrieved in the knowledge graph through the cosine similarity matching method; Convert policy parameters into physical control signals executable by the device; Sending instructions to the plasma generator execution unit through the industrial bus protocol; Combined with sensor feedback, control parameters are corrected in real time to form a closed-loop optimization.

8. The artificial intelligence-based plasma air purifier control method according to claim 7, characterized in that: The strategy parameters are converted into physical control signals executable by the device, specifically using a PID control algorithm to control the voltage of the plasma generator by adjusting the pulse width modulation PWM signal. The calculation formula is: PWM v =K p ·(V opt -V real )+K i ·∫(V opt -V real )dt Among them, PWM v K is a pulse width modulation (PWM) signal used to control the voltage of the plasma generator; p is the proportionality coefficient, K i is the integral coefficient; V opt is the optimal voltage value, V real It is the actual measured voltage value of the plasma generator.

9. The artificial intelligence-based plasma air purifier control method according to claim 1, characterized in that: The iterative detection of the pollution spatial distribution status dynamically corrects the reinforcement learning strategy through the difference analysis between the twin data and the real data, specifically: Use sensors to continuously collect real data on the spatial distribution of pollution in and around the plasma air purifier; Based on the established pollution feature knowledge base, system model and previously accumulated data, digital twin technology is used to generate twin data corresponding to the real scene; The collected real data and generated twin data are normalized, and the mean square error Euclidean distance indicator is used to calculate the difference between the twin data and the real data; According to the dimension with significant difference, the reward function in reinforcement learning is adjusted. Assuming that the original reward function is R(s,a), the adjustment factor based on the difference is α, and the adjusted reward function is R,(s,a), then: R,(s,a)=αR(s,a), where a is the action and s is the state; Using reinforcement learning algorithms such as the policy gradient algorithm, the policy network is updated using the adjusted reward function. The update formula for the policy network parameter θ is: i t+1 =θ t +η▽ θ logπ θ (a|s)R`(s,a) Among them, θ t is the parameter of the current policy network, η is the learning rate, π θ (a|s) is the probability of taking action a in state s; Integrate the newly collected real data and calculated difference information into the pollution characteristic knowledge base, and update and optimize the pollution data distribution pattern and working parameter characteristics; Based on the results of strategy correction, the parameters of the graph neural network model, reinforcement learning model, and simulation model are fine-tuned.

10. The plasma air purifier control system based on artificial intelligence is characterized by: include: The pollution feature knowledge base construction module collects the working parameters of the plasma air purifier system, extracts the pollution spatial distribution characteristics in the plasma air purifier through the improved PointNet++ network, and constructs the pollution feature knowledge base; A graph neural network model building module is used to connect the pollution feature knowledge base to the graph neural network model, use the graph neural network model to model and train the collected working parameters and the system's filtration efficiency for aerosols, and obtain a graph neural network model that can accurately predict filtration efficiency; The strategy library building module is used to optimize the system's working parameters by using a reinforcement learning algorithm combined with a trained graph neural network model to form a strategy library of working parameter combinations that maximize filtering efficiency; A knowledge graph construction module is used to use multi-physics field coupling simulation software to simulate and analyze the plasma air purification system, associate the simulation results with the reinforcement learning strategy library, and generate a parameter optimization knowledge graph; A control instruction generation module is used to generate plasma air purifier control instructions according to the knowledge graph to control the working state of the plasma air purifier; The strategy optimization module is used to iteratively detect the spatial distribution of pollution and dynamically correct the reinforcement learning strategy through difference analysis between twin data and real data.

Images