Air quality monitoring and purification regulation system and method based on sensor network

Abstract

The application discloses an air quality monitoring and purification control system and method based on a sensor network, and relates to the technical field of environmental data monitoring and processing and intelligent control; the application collects environmental data through a heterogeneous sensor network, and cleanses abnormal values; a dynamic air flow field directed weighted graph is constructed based on the time sequence correlation between sensor nodes, and a global air quality thermal map and a pollution source positioning are reconstructed by using a flow field corrected Kriging algorithm; a prediction model is constructed by introducing a graph neural network, and future air quality trends are predicted in combination with historical and real-time data; based on the prediction result, a dynamic cost function containing multiple targets of quality, energy consumption and noise is solved by using a quadratic programming algorithm, and an optimal control sequence is obtained; fresh air, purifiers and doors and windows are cooperatively dispatched to implement strategies such as positive pressure isolation or negative pressure pollution discharge; the application solves the problems of blind areas, control response lag and lack of cooperation of equipment in traditional monitoring, and realizes active and comfortable air quality management based on prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of environmental data monitoring, processing and intelligent control, and specifically, particularly relates to an air quality monitoring and purification regulation system and method based on a sensor network. Background Art

[0002] With the acceleration of urbanization and the improvement of residents' health awareness, indoor air quality has become an important indicator for measuring quality of life. Modern smart home systems are gradually transforming from single home appliance control to in-depth perception and active management of environmental ecology. People are no longer satisfied with simple haze removal, but put forward higher requirements for the comfort (low noise), energy saving of indoor environment and refined treatment for specific pollution scenarios (such as smoking, decoration formaldehyde, outdoor sand and dust), and particularly hope that the system has prediction capability instead of after-the-fact treatment.

[0003] Existing air purification solutions mainly rely on independent air purifiers or fresh air systems. Their typical working logic is to perform single-point monitoring based on sensors carried by the body, and when the concentration of monitored particulate matter (PM2.5) or carbon dioxide exceeds a preset threshold, the fan is passively turned on for filtration. Some high-end products have begun to try multi-point monitoring, using split detectors to link the main unit for operation. In terms of scenario application, fixed operation modes such as "automatic mode", "sleep mode" or "powerful mode" are mostly adopted, and devices usually operate independently and lack a linkage mechanism.

[0004] It can be seen from this that the prior art has the following defects: 1. Perception blind areas and attribution absence: single-point monitoring cannot reflect the air quality in room corners or actual activity areas of personnel, and cannot judge whether the pollution source comes from outdoor penetration or is generated indoors, leading to wrong treatment direction; 2. Lack of prediction and initiative: traditional control performs passive response based on current concentration threshold, which has obvious hysteresis, and cannot intercept pollution at the initial stage of diffusion, resulting in high-power purification only after pollution spreads to the whole house, with low efficiency; 3. Poor balance between comfort and efficiency: equipment often operates at full speed for rapid purification, generating huge noise, and cannot dynamically adjust the purification strategy according to the user's sleep or activity state, which seriously affects user experience. Summary of the Invention

[0005] (1) Technical problem to be solved Aiming at the problems in the related art, the present invention provides an air quality monitoring and purification regulation system and method based on a sensor network, so as to overcome the above-mentioned technical problems existing in the existing related art.

[0006] (2) Technical solution To solve the above technical problems, the present invention is implemented through the following technical solutions: S1. Collect air quality physicochemical index data, environmental physical field data, and human activity status characteristic data in the target space respectively, and perform outlier removal and time-series synchronization preprocessing through the optimized isolated forest algorithm to obtain a spatiotemporal multidimensional dataset. S2. Based on the spatiotemporal multidimensional dataset, calculate the time-series cross-correlation function and transmission delay between each sensor node, and construct a directed weighted graph of the dynamic airflow field; convert the directed weighted graph into a flow field vector matrix; S3. Based on the flow field vector matrix, the Kriging algorithm is modified by anisotropy correction to obtain the modified Kriging algorithm; the modified Kriging algorithm is used to reconstruct the virtual field of the blind area in the target space to obtain the real-time global air mass concentration gradient thermodynamic distribution model. S4. Input the real-time global air quality concentration gradient thermal distribution model into the trained graph neural network to obtain predicted pollution data; S5. Using the quadratic programming algorithm, solve the dynamic cost function based on the predicted pollution data to obtain the optimal control sequence; S6. Based on the optimal control sequence and the pollution source coordinates and outdoor environmental data located in S4, implement a differentiated airflow organization strategy; This invention achieves blind-spot-free visualization of indoor air quality and precise location of pollution sources through a heterogeneous sensor network and flow field correction Kriging algorithm; by introducing graph neural networks to deeply mine spatiotemporal data features, it transforms monitoring the present into predicting the future, enabling early prediction of pollution diffusion trends; based on the predicted data, it uses a quadratic programming algorithm to find the optimal solution between purification efficiency, energy consumption, and noise; and through differentiated airflow organization strategies, it achieves physical barrier and rapid sewage discharge; effectively solving the problems of traditional control lag and passive response, and realizing prediction-based proactive and collaborative environmental management.

[0007] Preferably, step S1 includes the following steps: S11. Deploy master nodes and slave nodes in the ground, breathing zone and ceiling areas of the target space to collect air quality physicochemical index data, environmental physical field data and personnel activity status characteristic data in the target space. At the same time, use millimeter-wave radar to collect personnel movement trajectory and breathing micro-motion signals to obtain an initial multi-source heterogeneous dataset. S12. The optimized isolated forest algorithm is used to detect outliers in the initial multi-source heterogeneous data and remove noise caused by sensor failure to obtain denoised multi-source heterogeneous data. S13. Establish a unified time reference axis, resample and align the denoised multi-source heterogeneous data with different sampling frequencies to form a spatiotemporal multidimensional dataset containing timestamps, three-dimensional spatial coordinates, pollutant concentration vectors and environmental parameter vectors. The above embodiments, through multi-layered hardware deployment and refined data preprocessing, construct a high-fidelity, noise-free environmental perception base, providing reliable input for the subsequent modeling of the invisible airflow field in S2.

[0008] Preferably, step S2 includes the following steps: S21. Calculate any two nodes in the sensor node set. i and j The Pearson correlation coefficient of the pollutant concentration sequence within the sliding time window; if the Pearson correlation coefficient is greater than the connectivity threshold, it is determined that there is an airflow connectivity edge between the nodes; S22. For node pairs with connected edges in the sensor node set, calculate the cross-correlation function of the node pairs, find the time delay that maximizes the cross-correlation function, and obtain the first time delay; if the first time delay > 0, define the airflow direction as... i point to j The weight of the edge is set to the reciprocal of the normalization constant and the absolute value of the first time delay to obtain the airflow transmission rate; S23. Define the set of sensor nodes as the vertex set of a directed weighted graph of the airflow field, and the airflow paths between nodes as the edge set; Traverse all node pairs in the sensor node set to construct a directed weighted graph of the dynamic airflow field; map the directed weighted graph of the dynamic airflow field into a flow field vector matrix; This invention uses the temporal correlation of sensor data to infer the physical flow field, solving the problem that indoor environments usually lack expensive anemometer arrays. By constructing a directed weighted graph and converting it into a vector matrix, the system can see the invisible air flow path. This not only provides anisotropic drift parameters for blind zone interpolation in S3, but also provides a physical basis for determining whether the pollution source is drifting with the wind or spreading against the wind in S5.

[0009] Preferably, step S3 includes the following steps: S31. Obtain the geometric boundary information of the target space and divide the target space into a voxel mesh; S32. Introduce the ordinary Kriging interpolation algorithm and define a semi-variogram function to describe spatial autocorrelation; combine the flow field vector matrix obtained in S2, introduce a drift term to perform anisotropic correction on the Euclidean distance, and obtain the modified Kriging algorithm. The anisotropy correction specifically involves reducing the attenuation rate of the interpolation weights in the upstream direction of the airflow and increasing the attenuation rate of the interpolation weights in the direction perpendicular to the airflow. S33. Using the modified Kriging algorithm, calculate the virtual concentration values ​​at locations where no sensors are deployed in the voxel grid, and generate a real-time global air quality concentration gradient thermal distribution model map. S34. Calculate the gradient vector of the concentration scalar field in the real-time global air quality concentration gradient thermodynamic distribution model diagram; perform an iterative search along the positive gradient direction until a local maximum point is found, and mark the coordinates of this point as the pollution source coordinates; This invention constructs a global air quality concentration gradient thermal distribution model by dividing the target space into a voxel grid and combining airflow field information to perform anisotropic correction on the Kriging interpolation algorithm. This method can intelligently adjust the spatial interpolation weights according to the airflow direction and use the gradient ascent algorithm to accurately locate the coordinates of pollution sources. Thus, under the condition of limited sensor deployment, it realizes high-precision virtual perception of the pollution status of blind spots in the indoor environment, providing reliable spatial situation information for subsequent intelligent ventilation control.

[0010] Preferably, step S4 includes the following steps: S41. Collect historical full-area air quality concentration gradient thermal distribution model maps, and collect the pollution source coordinate data and pollution concentration data corresponding to the historical full-area air quality concentration gradient thermal distribution model maps as label data to obtain training data; S42. Construct a graph neural network (GNN); train the graph neural network using training data, adjust the parameters of the graph neural network during the training process, and obtain the trained graph neural network. S43. Input the real-time global air quality concentration gradient thermal distribution model obtained in S3 into the trained graph neural network to obtain predicted pollution data; the predicted pollution data includes predicted pollution source coordinate data and predicted pollution concentration data in the target space; This step introduces machine learning algorithms, utilizing the powerful nonlinear fitting capabilities of graph neural networks. It not only considers the spatial pollutant concentration gradient but also incorporates dynamic flow field information in the flow field vector matrix, enabling it to accurately capture the diffusion patterns of pollutants in complex indoor environments. Compared to traditional physical deduction or linear extrapolation, the prediction model based on graph neural networks has higher robustness and accuracy, providing reliable data support for subsequent early control.

[0011] Preferably, step S5 includes the following steps: S51. Based on the predicted pollution data, define an initial dynamic cost function; the dynamic cost function is composed of a weighted sum of air quality deviation, equipment energy consumption, noise interference, and filter element wear. S52. Quantify the user activity status features extracted by S1 to obtain the user activity index; dynamically adjust the weight coefficients in the initial dynamic cost function based on the user activity index to obtain the real-time dynamic cost function. The rules for dynamically adjusting the weight coefficients in the initial dynamic cost function based on user activity index in S52 include: setting a threshold for human movement speed in sleep state and a threshold for movement speed in high-frequency activity state. When the user activity index is greater than or equal to the high-frequency activity state movement speed threshold, the weight coefficients in the initial dynamic cost function are adjusted according to the high-frequency activity state rules; when the human movement speed threshold in the sleep state is less than or equal to the user activity index and less than the high-frequency activity state movement speed threshold, the weight coefficients remain unchanged; when the user activity index is less than or equal to the human movement speed threshold in the sleep state, the weight coefficients in the initial dynamic cost function are adjusted according to the sleep state rules. S53. Set equipment operation constraints and state constraints to obtain constraint conditions; based on the constraint conditions, use a quadratic programming algorithm to solve for the control sequence that minimizes the real-time dynamic cost function to obtain the optimal control sequence; This invention transforms complex air pollution control decisions into a mathematical multi-objective dynamic optimization problem. By substituting the predicted data obtained from S4 into the cost function, the system is no longer based on the passive logic of turning on only when it is dirty, but on the proactive logic of intervening in advance when it is about to get dirty. At the same time, the quadratic programming algorithm ensures that a globally optimal equipment control strategy is obtained while meeting user comfort (low noise) and energy saving requirements.

[0012] Preferably, the step S53, which uses a quadratic programming algorithm to solve for the control sequence that minimizes the real-time dynamic cost function, includes the following steps: S531. Transform the real-time dynamic cost function J into the standard quadratic programming form to obtain the quadratic programming real-time dynamic cost function. S532. Transform the equipment operation constraints and state constraints into linear inequality constraint forms; S533. Use the interior point method or the effective set method to perform an iterative search in the feasible region until the KKT conditions are met, and output the optimal control sequence. This invention transforms model predictive control into a standard quadratic programming problem. Through constraint linearization and iterative solution, it ensures that a globally optimal control sequence is obtained while satisfying equipment and state constraints. This method guarantees the theoretical optimality and real-time computation of control decisions, and realizes multi-objective dynamic optimization control.

[0013] Preferably, step S6 includes the following steps: S61. Analyze the first step of the optimal control sequence output by S5 and decompose it into the intake fan speed, exhaust fan speed, air purifier fan speed and electric window opening. S62. Based on the pollution source coordinates located in S4 and the outdoor air quality data, select an airflow organization strategy and obtain airflow organization strategy instructions; if the pollution source is located indoors and the outdoor air quality is good, implement a negative pressure exhaust strategy, increase the exhaust fan speed and open the windows; if the pollution source is located outdoors, implement a positive pressure blocking strategy, close the windows and set the intake fan speed to be greater than the exhaust fan speed; if the pollution source is located in a local area indoors, implement a directional airflow strategy, adjust the angle of the air guide plate of the adjacent air purifier to point towards the pollution source and run it at full speed. S63. Issue airflow organization strategy instructions to each implementing agency; This invention intelligently switches between positive pressure blocking and negative pressure sewage discharge airflow organization strategies based on the nature of the pollution source (internal or external), cutting off the pollution path at the physical level. The efficiency is far higher than that of internal circulation purification of a single device. This collaborative game of multiple devices (fresh air, windows, purifier) ​​changes the inefficient situation of traditional independent cleaning of devices.

[0014] The sensor network-based air quality monitoring and purification control system is used to implement the above-mentioned sensor network-based air quality monitoring and purification control method, including: a sensing network and preprocessing module, a spatiotemporal modeling and deduction module, an intelligent prediction and decision-making module, and a collaborative execution module. Perception Network and Preprocessing Module: Corresponding to S1, it is responsible for driving heterogeneous sensor nodes and millimeter-wave radar, performing isolated forest anomaly detection and data spatiotemporal alignment, and outputting spatiotemporal multidimensional datasets; Spatiotemporal modeling and extrapolation module: corresponding to S2 and S3, responsible for constructing a directional weighted graph of dynamic airflow field and reconstructing global blind zone data using the flow field correction Kriging algorithm; Intelligent prediction and decision-making module: corresponding to S4 and S5, with a built-in pre-trained graph neural network model, responsible for predicting future pollution trends based on real-time data, and running a quadratic programming algorithm to generate the optimal control sequence including wind speed, mode and on / off state. Collaborative Execution Module: Corresponding to S6, it is responsible for parsing the control sequence and coordinating the scheduling of fresh air, air purifiers, and electric doors and windows through IoT protocols to execute differentiated airflow organization strategies.

[0015] (III) Beneficial Effects The present invention has the following beneficial effects: This invention achieves high-fidelity monitoring of the indoor environment by constructing a heterogeneous sensor network and combining it with the isolated forest algorithm. By utilizing the temporal correlation of sensor data to construct a directed weighted graph of dynamic airflow field, and combining it with anisotropic-corrected Kriging interpolation algorithm, it successfully realizes the virtual reconstruction of the whole-house air quality field under low-cost hardware, eliminating monitoring blind spots. At the same time, the gradient tracking-based algorithm can accurately locate the coordinates of pollution sources, enabling the system to distinguish between endogenous pollution (such as smoking) and exogenous pollution (such as seepage through cracks), providing reliable spatial situation information for subsequent precise treatment.

[0016] This invention breaks through the traditional "monitoring-feedback" lag control mode and introduces a graph neural network to construct an air quality prediction model. By combining historical data with real-time flow field characteristics, it can predict the trend of air quality changes in future periods. This gives the system the ability to predict and adjust the operating parameters (wind speed, mode, start and stop) of the purification equipment in advance before pollutants spread on a large scale, realizing a leap from passive governance to active governance and significantly improving governance efficiency.

[0017] This invention employs a multi-objective optimization strategy based on quadratic programming, which substitutes the predicted pollution data into a dynamic cost function that includes air quality, energy consumption, noise, and filter wear to solve for the optimal control sequence. The system can not only automatically adjust the weights according to the activity level of people (such as sleep and exercise) to balance comfort and purification efficiency, but also coordinate the scheduling of fresh air, air purifiers, and windows to execute differentiated airflow organization strategies (positive pressure blocking and negative pressure sewage discharge), realizing deep collaboration of multiple devices and people-oriented refined management.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a schematic flowchart of the air quality monitoring and purification control method based on sensor networks of the present invention. Figure 2 This is a schematic diagram of the air quality monitoring and purification control system based on sensor networks according to the present invention. Figure 3 This is a schematic diagram of the process for obtaining a real-time global air quality concentration gradient thermodynamic distribution model map in the air quality monitoring and purification control method based on sensor networks of the present invention. Figure 4 This is a schematic diagram of the process for locating the coordinates of pollution sources in the air quality monitoring and purification control method based on sensor networks of the present invention. Figure 5 This is a schematic diagram of the process for obtaining the optimal parameters of a graph neural network in the air quality monitoring and purification control method based on sensor networks of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0022] Please see Figure 1 This invention discloses an air quality monitoring and purification control method based on sensor networks, comprising the following steps: S1. Collect air quality physicochemical index data, environmental physical field data, and human activity status characteristic data in the target space respectively, and perform outlier removal and time-series synchronization preprocessing through the optimized isolated forest algorithm to obtain a spatiotemporal multidimensional dataset. S1 includes the following steps: S11. Deploy master nodes and slave nodes in the ground, breathing zone, and ceiling areas of the target space to collect air quality physicochemical index data, environmental physical field data, and personnel activity status characteristic data within the target space. The master nodes are equipped with laser particulate matter sensors, NDIR carbon dioxide sensors, and MEMS wind speed sensors; the slave nodes are equipped with temperature and humidity sensors and VOCs sensors. Simultaneously, use millimeter-wave radar to collect personnel movement trajectories and respiratory micro-motion signals to obtain an initial multi-source heterogeneous dataset. In specific implementation, the above embodiment S11 is as follows: This embodiment is deployed in a 120-square-meter three-bedroom, two-living-room house; main node deployment: one main node is deployed under the coffee table in the living room (0.5m from the ground), one at the head of the bed in the master bedroom (1.0m from the ground), and one at the desk in the children's room (1.0m from the ground); the main node hardware uses an STM32F4 microcontroller, integrating a Pantene G7 PM2.5 sensor (response time <1s, resolution 1ug / m3) and a SenseAir S8 sensor. CO2 sensor (accuracy ±30ppm) and high-sensitivity MEMS hot-film anemometer (for capturing minute airflow changes, range 0-10m / s); Slave node deployment: one slave node is deployed at each of the four corners of the ceiling in each room (2.8m from the ground), at the kitchen entrance, and at the bathroom entrance; the slave nodes use low-power ZigBee modules, integrating DHT22 temperature and humidity sensors and MEMS metal oxide VOCs sensors (mainly monitoring formaldehyde and benzene series compounds); Personnel detection radar deployment: a 60GHz millimeter-wave radar is deployed above the TV wall in the living room and at the entrance of each bedroom. This radar can output the target's XYZ spatial coordinates, movement speed (resolution 0.05m / s), and respiratory rate (for determining sleep status); Communication networking: all nodes form a mesh network via the ZigBee 3.0 protocol, converging to a central gateway; S12. The optimized isolated forest algorithm is used to detect outliers in the initial multi-source heterogeneous data and remove noise caused by sensor failure to obtain denoised multi-source heterogeneous data. In specific implementation, the above embodiment S12 is as follows: During initial operation or during off-peak hours in the early morning, the system constructs M1=100 isolated trees using historical data from the most recent 10 minutes (n1=600 points); when a newly collected PM2.5 value of 999 enters the system, due to its significant difference from the historical normal distribution, it is cut and separated near the root node of the tree, resulting in an average path length E(h(x)). t Extremely short; according to the formula ;in, s c Indicates abnormal scores. c ( n 1) This means that the average path length normalization factor calculated for a given sample size n1 results in an anomaly score of 0.92 (greater than the threshold of 0.85). The system determines this to be sensor electrical noise and replaces the noise point with the median of the five most recent valid data points in the sliding window. S13. Establish a unified time reference axis, resample and align the denoised multi-source heterogeneous data with different sampling frequencies to form a spatiotemporal multidimensional dataset containing timestamps, three-dimensional spatial coordinates, pollutant concentration vectors and environmental parameter vectors. In specific implementation, the above embodiment S13 is as follows: Since the master node sampling rate is 1Hz, the slave node sampling rate is 0.2Hz (power-saving mode), and the radar data sampling rate is 10Hz; this embodiment sets a unified time reference axis with a step size of 1 second; for slave node data, linear interpolation is used to fill in missing time values; for radar data, downsampling and averaging are used to align to 1Hz; the final generated spatiotemporal multidimensional dataset D has each row of data structure defined as: Dt={timestamp, node ID, (spatial coordinates x, y, z), [PM2.5 concentration, carbon dioxide concentration, volatile organic compound concentration], [temperature, humidity, wind speed vector], [personnel movement speed, personnel position coordinates]}; in this embodiment, Dt={2026-01-01 09:00:01, Node A, (2.5, 3.0, 0.5), [35, 600, 0.02], [24.5, 50, (0.1, 0.0, 0)], [0.05, (2.5, 3.0)]}; The above embodiments achieve high-precision, multi-dimensional perception of the indoor environment by constructing a hierarchical heterogeneous sensor network, combining the human perception capability of millimeter-wave radar, and using isolated forests for data cleaning; overcoming the shortcomings of traditional single sensors, such as limited monitoring range and susceptibility to interference, and providing a reliable data foundation for subsequent flow field modeling and precise control. S2. Based on the spatiotemporal multidimensional dataset, calculate the time-series cross-correlation function and transmission delay between each sensor node, and construct a directed weighted graph of the dynamic airflow field; convert the directed weighted graph into a flow field vector matrix; S2 includes the following steps: S21. Calculate any two nodes in the sensor node set. i and j The Pearson correlation coefficient of the pollutant concentration sequence within the sliding time window; if the Pearson correlation coefficient is greater than the connectivity threshold, it is determined that there is an airflow connectivity edge between the nodes; In specific implementation, the above embodiment S22 specifically involves: extracting the PM2.5 concentration sequence from the spatiotemporal multidimensional dataset; and setting the sliding window length. W h =60s; In this embodiment, the PM2.5 concentration sequence of node A (located at the entrance) is as follows: X A ,node B The PM2.5 concentration sequence (located in the center of the living room) is as follows: X B Calculate the Pearson correlation coefficient. ;in r AB Represents node A and node B The Pearson correlation coefficient between them; s A , s B Representing the concentration sequence X A and X B The standard deviation; this embodiment sets the connectivity threshold. b link =0.75, calculated as follows r AB =0.88> b link =0.75 indicates that the concentration changes at the two points are synchronized, suggesting a direct airflow channel (without physical obstruction) between them. A connecting edge is then established in the directed weighted graph of the airflow field. E AB ;like r AC =0.2, indicating that the node A With nodes C If the airflow between them is obstructed or the distance is too great, no connecting edge should be established; S22. For node pairs with connected edges in the sensor node set, calculate the cross-correlation function of the node pairs, find the time delay that maximizes the cross-correlation function, and obtain the first time delay; if the first time delay > 0, define the airflow direction as... i point to j The weight of the edge is set to the reciprocal of the normalization constant and the absolute value of the first time delay to obtain the airflow transmission rate; In specific implementation, the above embodiment S23 specifically refers to: for nodes with established connections... A and B Calculate the cross-correlation function. R AB ( t )=E[ X A (t)* X B ( t + t )];in, R AB ( t ) represents a node A With nodes B The cross-correlation function between the concentration sequences, i.e., the correlation between the two sequences after a delay. t The degree of similarity; E[] represents the expectation operator, indicating averaging over time; in the lag time t The search for the peak value is within the range of [−30s, 30s]; the calculation results show that when the lag time τ = +10s, R AB Reaching the maximum value; at this point, it means the node...A The concentration at that location was transmitted to the node after a 10-second delay. B Therefore, the current airflow direction is determined to be from node A to B; the weights of the edges are then calculated. w AB = K / | τ|=10 / 10=1.0, where K The normalization constant is set to 10 in this embodiment; a larger weight indicates a smaller transmission delay, i.e. a faster wind speed; if τ is negative, the airflow direction is opposite; in this embodiment, it is set to 10 to map the reciprocal of the time delay to a numerically stable weight range (0.1-10.0), to avoid the weight from tending to infinity when the delay τ is extremely small (the wind speed is extremely fast), which would cause matrix operation overflow, and at the same time, it also unifies the magnitude of the flow field vector. S23. Define the set of sensor nodes as the vertex set of a directed weighted graph of the airflow field, and the airflow paths between nodes as the edge set; Traverse all node pairs in the sensor node set to construct a directed weighted graph of the dynamic airflow field; map the directed weighted graph of the dynamic airflow field into a flow field vector matrix; In specific implementation, the above embodiment S24 specifically involves: traversing all node pairs to generate a directed weighted graph of the dynamic airflow field. G =( V , E , W ),in, V Represents the set of vertices. E Denotes the set of edges. W Represents the set of edge weights; based on the coordinates of the nodes in physical space ( x , y , z ), the diagram G This is transformed into a flow field vector matrix; specifically: for each node in space... i Iterate through all its in-degree neighbors to obtain the unit direction vector; multiply this unit vector by the corresponding edge weight. w ji (i.e., transmission rate), to obtain the component vector; perform vector synthesis (vector addition) on all neighboring component vectors to obtain the node. i The resultant flow field vector at that location is used to represent the net direction and intensity of the airflow at that location; For any node in the flow field vector matrix space i Its flow field vector is equal to that of all its connected neighboring nodes. j Weighted vector sum: ;in, V i This represents the flow field vector at the i-th node in the flow field vector matrix; P i nodei The position vector, P i =( x i , y i , z i ), i.e., sensor i Coordinates in three-dimensional space; P j node j The position vector, P j =( x j , y j , z j ); No ( i ) represents the nodes in G. i There exists a set of nodes directly connected by directed edges; w ji Indicates from node j Pointing to node i The weights of the edges; the flow field vector matrix intuitively describes the flow direction and velocity of pollutants at each sensor node at the current moment, providing key drift parameters for the interpolation correction of S3; The above embodiment S2 innovatively constructs a dynamic airflow model based on graph theory by analyzing the correlation and lag of multi-node time series data. This method does not require expensive CFD (computational fluid dynamics) real-time simulation. It can infer the topology of airflow based solely on sensor data and accurately identify the dynamic process of "pollutants drifting from the door to the sofa", providing key directional guidance for subsequent active interception. S3. Based on the flow field vector matrix, the Kriging algorithm is modified by anisotropy correction to obtain the modified Kriging algorithm; the modified Kriging algorithm is used to reconstruct the virtual field of the blind area in the target space to obtain the real-time global air quality concentration gradient thermodynamic distribution model map, and the pollution source coordinates are located by reverse tracking based on the gradient ascent algorithm. Please see Figure 3 S3 includes the following steps: S31. Obtain the geometric boundary information of the target space and divide the target space into a voxel mesh; In specific implementation, the above embodiment S31 specifically involves: using LiDAR scanning or a pre-set BIM model to establish a 3D coordinate system for the room, dividing the 120 square meter space with a floor height of 2.8 meters into a 0.5m×0.5m×0.5m voxel grid, totaling 2688 grid points; marking inaccessible areas such as walls and furniture; S32. Introduce the ordinary Kriging interpolation algorithm and define a semi-variogram function to describe spatial autocorrelation; combine the flow field vector matrix obtained in S2, introduce a drift term to perform anisotropic correction on the Euclidean distance, and obtain the modified Kriging algorithm. The anisotropy correction specifically involves reducing the attenuation rate of the interpolation weights in the upstream direction of the airflow and increasing the attenuation rate of the interpolation weights in the direction perpendicular to the airflow. In specific implementation, the above embodiment S32 is as follows: Traditional Kriging interpolation assumes spatial isotropy, meaning the closer the distance, the stronger the correlation; however, in airflow scenarios, the impact of a pollution source farther upwind on the current point may be greater than the impact of a nearby source in the crosswind; this embodiment uses an anisotropic distance formula to replace Euclidean distance. In the downwind direction, the "statistical distance" is compressed, increasing the interpolation weight; in the vertical wind direction, the "statistical distance" is stretched, decreasing the weight; this solves the data errors caused by the traditional Kriging interpolation method; the specific formula for the corrected Kriging algorithm is as follows: ;in, h ij Point i and points j The corrected distance between them (i.e., the anisotropic distance) is used to replace the Euclidean distance in calculating the semivariogram value; x 1 i , x 1 j Representing points respectively i and points j Coordinate vectors in three-dimensional space, i.e. ( x j , y j , z j ) T , I Represents the identity matrix. β This represents the anisotropy coefficient, which is set to 2.0 in this example. It is used to control the intensity of the airflow's distance correction. β The larger the value, the more significant the impact of airflow direction on distance measurement; β When = 0, the corrected distance is equivalent to the Euclidean distance; V avg Point i and points j The average airflow velocity vector in the region between; V avg V T avg This represents the cross product of the average velocity vector and its transpose. S33. Using the modified Kriging algorithm, calculate the virtual concentration values ​​at locations where no sensors are deployed in the voxel grid, and generate a global air quality concentration gradient thermodynamic distribution model. In specific implementation, the above embodiment S33 specifically refers to: for the coordinates of the room corner (blind spot) P blind The system selects the five closest sensor nodes and constructs the Kriging equations using the distance formula modified by S32: The Kriging equations consist of two parts: dl + uα = c , α T λ= 1; among which, d Describes a symmetric matrix, containing h ij ; α This represents a 5×1 column vector consisting entirely of 1s. α =[1,1,1,1,1] T ; α T express α transpose , even A 5×1 row vector consisting entirely of 1s; c This represents a 5×1 column vector containing known nodes. i With blind spot P blind The semivariogram values ​​between; l Let λ represent a 5×1 column vector containing the Kriging weights to be determined, i.e., λ = [λ1, λ2, ..., λ5]. T These weights will be used to weight the known concentration values; u Let λ represent the scalar Lagrange multipliers; calculate the Kriging equations to obtain the weight vector λ. Based on Kriging weights l Calculate virtual concentration Among them, the distance to the blind spot P blind The locations of the five most recent sensor nodes are denoted as follows: x 2 i , (i=1,…,5); the corresponding concentration measurement value is Z( x 2 i ); l i Indicates the first iKriging weights; by calculating all voxel grids, a real-time global air quality concentration gradient thermal distribution model Map3D containing the PM2.5 concentration values ​​of the whole house is generated; for example, the model shows that although the sensor reading in the living room is 50, the virtual reading in the corner of the sofa is as high as 120 (presumably due to airflow accumulation). S34. Calculate the gradient vector of the concentration scalar field in the real-time global air quality concentration gradient thermodynamic distribution model diagram; perform an iterative search along the positive gradient direction until a local maximum point is found, and mark the coordinates of this point as the pollution source coordinates; In specific implementation, the above embodiment S34 specifically involves: calculating the gradient vector of each grid point in Map3D; the calculation formula is: ;in, These represent the concentrations at... x , y , z The partial derivatives in each direction represent the rate of change of concentration along each coordinate axis; The source tracing algorithm is used to find the coordinates of the main pollution source. Specifically, the sensor node with the highest current concentration in the whole house is selected as the starting point. P start Along the positive gradient direction ∇ C Move one step (0.5m); repeat until the gradient magnitude is |∇ C | Approaching 0, indicating a local maximum; the coordinates of the local maximum in this embodiment. P source Located at (8.5, 0.0, 1.5) (corresponding to the location of the balcony window gap), the pollution source is determined to be external infiltration; if the coordinates of the maximum point... P source (4.0, 3.5, 0.8) (corresponding to the coffee table position), the system determines that the pollution source is internal (such as smoking). This coordinate information will determine whether to open or close the window in S5, which is crucial; Please see Figure 4 Scenario demonstration: This embodiment detects an overall increase in concentration in the living room area; starting from any point, along ∇ C The movement follows the direction of maximum concentration increase, with the path sequentially from sofa to coffee table to the gap in the balcony sliding door; ultimately, the local maximum point is located at the balcony door gap (coordinates...). x =8.5, y =3.0, z =1.5), at this point the pollution source is determined to be outdoor infiltration (possibly due to severe outdoor smog and poorly sealed doors and windows). If the maximum value is located above the sofa, it is determined to be an internal source (such as smoking). The above embodiments successfully constructed a "virtual sensor" by introducing a flow field correction Kriging interpolation algorithm, achieving high-resolution air quality field reconstruction with limited hardware costs; this is a key prerequisite for subsequently obtaining predicted pollution data and formulating differentiated purification strategies. S4. Input the real-time global air quality concentration gradient thermal distribution model into the trained graph neural network to obtain predicted pollution data; S4 includes the following steps: S41. Collect historical full-area air quality concentration gradient thermal distribution model maps, and collect the pollution source coordinate data and pollution concentration data corresponding to the historical full-area air quality concentration gradient thermal distribution model maps as label data to obtain training data; S42. Construct a graph neural network (GNN); train the graph neural network using training data, adjust the parameters of the graph neural network during the training process, and obtain the trained graph neural network. In specific implementation, the above embodiment S42 is as follows: the network structure of the trained graph neural network is that the input layer receives the node features (concentration) and edge features (flow field vectors) of the voxel grid; the intermediate layer contains several graph convolutional layers (GCNLayer) for extracting spatial topological features, and a long short-term memory network layer (LSTM layer) for capturing the temporal evolution law; the output layer outputs the predicted heat map of future time through a fully connected layer; The parameters of the graph neural network (including the weight matrix and bias of the graph convolutional layer, the gating mechanism parameters of the long short-term memory network layer, and the weights and biases of the fully connected output layer) can be adjusted using various optimization algorithms (including particle swarm optimization, sparrow algorithm, and genetic algorithm). Please refer to [link to relevant documentation]. Figure 5 In this embodiment, the parameters of the graph neural network are optimized using the particle swarm optimization algorithm, specifically as follows: Construct a particle set, and set the size of the particle set to be... o The particle set is then represented as: ,in, h i This represents the i-th particle in the particle set; o The total number of particles is represented; the parameter search space and maximum number of training iterations of the graph neural network are defined; the fitness function is defined as the Euclidean distance formula between the predicted data and the labeled data in the training data of the graph neural network; Based on the parameter search space of a graph neural network, a set of initial particle positions is constructed. ,in, p i Represents the first particle in the set of particles i The initial position of each particle; the position of each particle in the initial position set is used as a parameter of a random graph neural network; An iterative update operation is performed on the particles in the particle set. The fitness value of each particle in the particle set is calculated according to the fitness function. The fitness of each particle in the particle set is updated from high to low fitness value. In each round of iteration, the best individual particle and the global best particle are obtained. During the update process, the positions of particles with the top 10% fitness are retained, and the middle positions of particles with the top 10%-50% fitness are retained (i.e., 40% of the positions enter the next round). The remaining 50% are obtained through random particle movement. The best individual particle is the particle with the highest fitness value in each round of iteration, and the global best particle is the particle with the highest fitness value in all rounds of iteration. Repeat the iteration until the maximum number of training iterations is reached, then stop iterating and use the globally optimal particle as the parameter of the graph neural network. S43. Input the real-time global air quality concentration gradient thermal distribution model obtained in S3 into the trained graph neural network to obtain predicted pollution data; the predicted pollution data includes predicted pollution source coordinate data and predicted pollution concentration data in the target space; S5. Using the quadratic programming algorithm, solve the dynamic cost function based on the predicted pollution data to obtain the optimal control sequence; S5 includes the following steps: S51. Based on the predicted pollution data, define an initial dynamic cost function; the dynamic cost function is composed of a weighted sum of air quality deviation, equipment energy consumption, noise interference, and filter element wear. S52. Quantify the user activity status features extracted by S1 to obtain the user activity index; dynamically adjust the weight coefficients in the initial dynamic cost function based on the user activity index to obtain the real-time dynamic cost function. In specific implementation, the above embodiments S41 and S42 specifically refer to: the initial dynamic cost function ; where 10 represents the 10 future time steps set; a 1. a 2. a 3. a 4 represents air quality deviation and equipment energy consumption, respectively. Ene Noise interference Noi and filter element wear Fil Weighting coefficients; C ta This indicates the target PM2.5 concentration (e.g., 35 μg / m³, corresponding to the excellent air quality standard). C pr Indicates the PM2.5 concentration at a future time; The rules for dynamically adjusting the weight coefficients in the initial dynamic cost function based on user activity metrics in S52 include: S521. Set the threshold for human movement speed during sleep and the threshold for movement speed during high-frequency activity. S522. When the user activity index is greater than or equal to the high-frequency activity state movement speed threshold, the weight coefficients in the initial dynamic cost function are adjusted using the high-frequency activity state rules; when the sleep state human movement speed threshold is less than or equal to the user activity index and less than the high-frequency activity state movement speed threshold, the weight coefficients remain unchanged; when the user activity index is less than or equal to the sleep state human movement speed threshold, the weight coefficients in the initial dynamic cost function are adjusted using the sleep state rules. Furthermore, dynamically adjust according to the scenario. a 1. a 2. a 3. a The value of 4; the sleep state rules are as follows: Scenario 1: Late at night, the millimeter-wave radar in S1 detects that the movement speed of the person in the master bedroom is Iact=0.05m / s<0.1m / s (stationary), and the breathing rate is stable. The system determines it to be "sleep mode"; set noise weight. a 3 = 10.0 (maximum), mass weight a 1=0.5 (allowing for a slow decrease in concentration, not pursuing extreme speed); at this time, the optimizer tends to choose a strategy of low wind speed and long-term operation to avoid noise waking users; Scenario 2: Smoking at a party; radar detects multiple people moving at high frequency in the living room, Iact=0.8m / s, and the heat map in S3 shows an extremely high concentration gradient at the coffee table (strong pollution source); forced a 1 = 10.0, noise weighting a 3 = 0.1 (At this point, the user is not sensitive to noise, but the health risk is high); at this point, the optimizer outputs "jet mode" and runs at full speed; S53. Set equipment operation constraints and state constraints to obtain constraint conditions; based on the constraint conditions, use a quadratic programming algorithm to solve for the control sequence that minimizes the real-time dynamic cost function. In specific implementation, the above embodiment S44 specifically includes the following constraints: 0 ≤ fan speed ≤ 100% (physical limitation of fan speed); C pr ≥0 (concentration is non-negative); | g ( k )− g (k−1)|≤ motor speed surge threshold (to prevent excessive sudden changes in motor speed from damaging the equipment); use the interior point method to solve the quadratic programming problem and obtain the optimal control sequence for the next 10 time steps (e.g., the next 10 minutes); The step S53, which uses a quadratic programming algorithm to solve for the control sequence that minimizes the real-time dynamic cost function, includes the following steps: S531. Transform the real-time dynamic cost function J into the standard quadratic programming (QP) form to obtain the quadratic programming real-time dynamic cost function. S532. Transform equipment operation constraints (such as fan speed 0-100%) and state constraints (concentration non-negative) into linear inequality constraint forms; S533. Use the interior point method or the effective set method to perform an iterative search in the feasible region until the KKT conditions are met, and output the optimal control sequence. In specific implementation, the above embodiment S53 is as follows: the real-time dynamic cost function of quadratic programming is... ;in, U This represents a control variable with a dimension of 10×5; H This represents the positive definite Hessian matrix composed of the coefficients of the quadratic terms in the real-time dynamic cost function; f The gradient vector is composed of the coefficients of the first-order terms in the cost function. The interior point method iteratively approximates the optimal solution from within the feasible region and is suitable for large-scale problems. The effective set method iteratively updates the effective constraint set on the boundary and is suitable for small- to medium-scale problems. Both can efficiently solve quadratic programming problems. In specific implementation, the choice can be made dynamically according to the size of the target space and the amount of data. In this embodiment, a 120-square-meter three-bedroom, two-living-room house is taken as an example, so the effective set method is chosen. Using the control variables (10 × 5 = 50 variables) for the next 10 time steps as the optimization object, the weighted terms in the cost function are expanded to construct a 50 × 50 positive definite Hessian matrix H; constraint matrix A cons This includes a limit on the rate of change of motor speed (e.g., a change of no more than 10% per second); using an embedded QP solver (OSQP in this embodiment), iterative calculations are completed within 200ms, outputting a sequence: 80% fresh air at t=1, 60% fresh air at t=2, ... 20% fresh air at t=10. The system extracts the command at time t=1 and issues it; the KKT conditions include the stationarity condition ∇ U L = HU + f + A cons T u =0, meaning the linear combination of the objective function gradient and the constraint gradient is zero, where, A cons T u Weighted sum of constrained gradients; original feasibility A cons U ≤ b cons That is, to find the solution UAll constraints (device limitations, state limitations, etc.) must be met. A cons Represents the constraint matrix. b cons Represents the constraint vector; duality feasibility u ≥0, meaning the Lagrange multipliers must be non-negative; complementary relaxation condition. u i ( A cons U - b cons ) i =0, meaning for each constraint i Or u i =0 (the constraint has no effect, i.e., it is relaxed), or ( A cons U - b cons ) i =0 (constraint is tight, equality holds). Only constraints that are active (tight constraints) correspond to non-zero multipliers; constraints that are inactive do not affect the optimal solution. u i Indicates about the first i A constrained Lagrange multiplier component; S6. Based on the optimal control sequence and the pollution source coordinates and outdoor environmental data located in S4, implement a differentiated airflow organization strategy; S6 includes the following steps: S61. Analyze the first step of the optimal control sequence output by S5 and decompose it into the intake fan speed, exhaust fan speed, air purifier fan speed and electric window opening. S62. Based on the pollution source coordinates located in S4 and the outdoor air quality data, select an airflow organization strategy and obtain airflow organization strategy instructions; if the pollution source is located indoors and the outdoor air quality is good, implement a negative pressure exhaust strategy, increase the exhaust fan speed and open the windows; if the pollution source is located outdoors, implement a positive pressure blocking strategy, close the windows and set the intake fan speed to be greater than the exhaust fan speed; if the pollution source is located in a local area indoors, implement a directional airflow strategy, adjust the angle of the air guide plate of the adjacent air purifier to point towards the pollution source and run it at full speed. In specific implementation, the above embodiment S62 specifically involves: based on the pollution source coordinates output in S4... P source Based on the outdoor PM2.5 data collected by S1, the airflow organization strategy is fine-tuned according to the analyzed control parameters; Strategy F1: Positive pressure blocking strategy (anti-haze), the criterion is... P sourceLocated in the window gap (S3 conclusion), and with outdoor PM2.5 > 150; the action to be taken is to forcibly close the electric window (opening = 0); control the fresh air intake speed Vin = 80% and the exhaust speed Vout = 40%; Vin > Vout causes the indoor air pressure to rise, forming a slight positive pressure (> 10 Pa), and the airflow overflows outward through the gap, physically preventing the infiltration of dirty outdoor air; Strategy F2: negative pressure exhaust strategy (smoke removal), the criterion is... P source Located on the indoor coffee table, and with outdoor PM2.5 < 35 (excellent air quality); the action to be taken is to open the electric window (opening degree = 50%), control the fresh air exhaust speed Vout = 100%, and the air intake speed Vin = 0%; creating indoor negative pressure, using clean outdoor air to rush in through the window, quickly squeezing the indoor smoke to the exhaust vent, which is much more efficient than simple internal circulation purification; Strategy F3: Directional airflow strategy (cooperative interception); the criterion is that the pollution source is in area A21 and the personnel are in area B21; the action to be taken is to adjust the angle of the air guide plate of the purifier in area A21 so that its air outlet direction is directed towards the entrance of area B21, forming an air curtain wall, using turbulence to block the diffusion of pollutants to the clean area; S63. Issue airflow organization strategy instructions to each implementing agency; The above embodiments demonstrate the high flexibility and closed-loop error correction capability of the present invention at the execution level; through the "positive pressure / negative pressure" switching logic, the system cleverly utilizes air pressure difference to solve different types of pollution problems; through continuous prediction, the system can adapt to dynamic changes in the environment (such as if someone suddenly lights a cigarette, it will predict that the pollutant concentration will rise, and thus implement corresponding execution strategies), ensuring that the purification effect is always in the optimal state; it changes the situation of traditional equipment working alone and builds an intelligent air ecosystem that is perceptive, intelligent, and collaborative.

[0023] For further details, please refer to Figure 2 An air quality monitoring and purification control system based on sensor networks is used to implement the above-mentioned air quality monitoring and purification control method based on sensor networks. It includes a sensing network and preprocessing module, a spatiotemporal modeling and deduction module, an intelligent prediction and decision-making module, and a collaborative execution module. The sensing network and preprocessing module: corresponding to S1, is responsible for driving heterogeneous sensor nodes and millimeter-wave radar, performing isolated forest anomaly detection and data spatiotemporal alignment, and outputting spatiotemporal multidimensional datasets; The spatiotemporal modeling and extrapolation module, corresponding to S2 and S3, is responsible for constructing a directional weighted graph of the dynamic airflow field and reconstructing the global blind zone data using the flow field correction Kriging algorithm. The intelligent prediction and decision-making module, corresponding to S4 and S5, has a built-in trained graph neural network model. It is responsible for predicting future pollution trends based on real-time data and running a quadratic programming algorithm to generate the optimal control sequence including wind speed, mode, and on / off state. The collaborative execution module, corresponding to S6, is responsible for parsing the control sequence and coordinating the scheduling of fresh air, air purifiers, and electric doors and windows through the Internet of Things protocol to execute differentiated airflow organization strategies.

[0024] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0025] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for air quality monitoring and purification control based on sensor networks, characterized in that, Includes the following steps: S1. Collect air quality physicochemical index data, environmental physical field data, and human activity status characteristic data in the target space respectively, and perform outlier removal and time-series synchronization preprocessing through the optimized isolated forest algorithm to obtain a spatiotemporal multidimensional dataset. S2. Based on the spatiotemporal multidimensional dataset, calculate the time-series cross-correlation function and transmission delay between each sensor node, and construct a directed weighted graph of the dynamic airflow field; convert the directed weighted graph into a flow field vector matrix; S3. Based on the flow field vector matrix, the Kriging algorithm is modified by anisotropy correction to obtain the modified Kriging algorithm; the modified Kriging algorithm is used to reconstruct the virtual field of the blind area in the target space to obtain the real-time global air mass concentration gradient thermodynamic distribution model. S4. Input the real-time global air quality concentration gradient thermal distribution model into the trained graph neural network to obtain predicted pollution data; S5. Using the quadratic programming algorithm, solve the dynamic cost function based on the predicted pollution data to obtain the optimal control sequence; S6. Implement differentiated airflow organization strategies based on predicted pollution data, optimal control sequences, and outdoor environmental data; S3 includes: S31. Obtain the geometric boundary information of the target space and divide the target space into a voxel mesh; S32. Introduce the ordinary Kriging interpolation algorithm and define a semi-variogram function to describe spatial autocorrelation; combine the flow field vector matrix obtained in S2, introduce a drift term to perform anisotropic correction on the Euclidean distance, and obtain the modified Kriging algorithm. The anisotropy correction specifically involves reducing the attenuation rate of the interpolation weights in the upstream direction of the airflow and increasing the attenuation rate of the interpolation weights in the direction perpendicular to the airflow. S33. Using the modified Kriging algorithm, calculate the virtual concentration values ​​at locations where no sensors are deployed in the voxel grid, and generate a real-time global air quality concentration gradient thermal distribution model map. S34. Calculate the gradient vector of the concentration scalar field in the real-time global air quality concentration gradient thermodynamic distribution model diagram; perform iterative search along the positive gradient direction until a local maximum point is found, and mark the coordinates of this point as the pollution source coordinates.

2. The air quality monitoring and purification control method based on sensor networks according to claim 1, characterized in that, S1 includes the following steps: S11. Deploy master nodes and slave nodes in the ground, breathing zone and ceiling areas of the target space to collect air quality physicochemical index data, environmental physical field data and personnel activity status characteristic data in the target space. At the same time, use millimeter-wave radar to collect personnel movement trajectory and breathing micro-motion signals to obtain an initial multi-source heterogeneous dataset. S12. The optimized isolated forest algorithm is used to detect outliers in the initial multi-source heterogeneous data and remove noise caused by sensor failure to obtain denoised multi-source heterogeneous data. S13. Establish a unified time reference axis, resample and align the denoised multi-source heterogeneous data with different sampling frequencies to form a spatiotemporal multidimensional dataset containing timestamps, three-dimensional spatial coordinates, pollutant concentration vectors and environmental parameter vectors.

3. The air quality monitoring and purification control method based on sensor networks according to claim 2, characterized in that, S2 includes the following steps: S21. Calculate any two nodes in the sensor node set. i and j The Pearson correlation coefficient of the pollutant concentration sequence within the sliding time window; if the Pearson correlation coefficient is greater than the connectivity threshold, it is determined that there is an airflow connectivity edge between the nodes; S22. For node pairs with connected edges in the sensor node set, calculate the cross-correlation function of the node pairs, find the time delay that maximizes the cross-correlation function, and obtain the first time delay; if the first time delay > 0, define the airflow direction as... i point to j The weight of the edge is set to the reciprocal of the normalization constant and the absolute value of the first time delay to obtain the airflow transmission rate; S23. Define the set of sensor nodes as the vertex set of a directed weighted graph of the airflow field, and the airflow paths between nodes as the edge set; Traverse all node pairs in the sensor node set to construct a directed weighted graph of the dynamic airflow field; map the directed weighted graph of the dynamic airflow field into a flow field vector matrix.

4. The air quality monitoring and purification control method based on sensor networks according to claim 1, characterized in that, S4 includes the following steps: S41. Collect historical full-area air quality concentration gradient thermal distribution model maps, and collect the pollution source coordinate data and pollution concentration data corresponding to the historical full-area air quality concentration gradient thermal distribution model maps as label data to obtain training data; S42. Construct a graph neural network; train the graph neural network using training data, adjust the parameters of the graph neural network during the training process, and obtain the trained graph neural network. S43. Input the real-time global air quality concentration gradient thermal distribution model obtained in S3 into the trained graph neural network to obtain predicted pollution data; the predicted pollution data includes predicted pollution source coordinate data and predicted pollution concentration data in the target space.

5. The air quality monitoring and purification control method based on sensor networks according to claim 4, characterized in that, S5 includes the following steps: S51. Based on the predicted pollution data, define an initial dynamic cost function; the dynamic cost function is composed of a weighted sum of air quality deviation, equipment energy consumption, noise interference, and filter element wear. S52. Quantify the user activity status features extracted by S1 to obtain the user activity index; dynamically adjust the weight coefficients in the initial dynamic cost function based on the user activity index to obtain the real-time dynamic cost function. The rules for dynamically adjusting the weight coefficients in the initial dynamic cost function based on user activity index in S52 include: setting a threshold for human movement speed in sleep state and a threshold for movement speed in high-frequency activity state. When the user activity index is greater than or equal to the high-frequency activity state movement speed threshold, the weight coefficients in the initial dynamic cost function are adjusted according to the high-frequency activity state rules; when the human movement speed threshold in the sleep state is less than or equal to the user activity index and less than the high-frequency activity state movement speed threshold, the weight coefficients remain unchanged; when the user activity index is less than or equal to the human movement speed threshold in the sleep state, the weight coefficients in the initial dynamic cost function are adjusted according to the sleep state rules. S53. Set equipment operation constraints and state constraints to obtain constraint conditions; based on the constraint conditions, use a quadratic programming algorithm to solve for the control sequence that minimizes the real-time dynamic cost function to obtain the optimal control sequence.

6. The air quality monitoring and purification control method based on sensor networks according to claim 5, characterized in that, The step S53, which uses a quadratic programming algorithm to solve for the control sequence that minimizes the real-time dynamic cost function, includes the following steps: S531. Transform the real-time dynamic cost function J into the standard quadratic programming form to obtain the quadratic programming real-time dynamic cost function. S532. Transform the equipment operation constraints and state constraints into linear inequality constraint forms. S533. Use the interior point method or the effective set method to perform an iterative search within the feasible region until the KKT conditions are met, and output the optimal control sequence.

7. The air quality monitoring and purification control method based on sensor networks according to claim 5, characterized in that, S6 includes the following steps: S61. Analyze the first step of the optimal control sequence output by S5 and decompose it into the intake fan speed, exhaust fan speed, air purifier fan speed and electric window opening. S62. Based on the pollution source coordinates located in S4 and the outdoor air quality data, select an airflow organization strategy and obtain an airflow organization strategy instruction; the airflow organization strategy instruction includes a negative pressure sewage discharge strategy, a positive pressure blocking strategy, and a directional flow guidance strategy. S63. Issue airflow organization strategy instructions to each implementing agency.

8. An air quality monitoring and purification control system based on sensor networks, characterized in that: The air quality monitoring and purification control method based on sensor networks as described in any one of claims 1-7 includes a sensing network and preprocessing module, a spatiotemporal modeling and deduction module, an intelligent prediction and decision-making module, and a collaborative execution module.

Images