Method and system for monitoring air quality in high and large space based on physical information neural network

Abstract

The present application relates to the field of artificial intelligence and environmental monitoring technology, and discloses a high and large space air quality monitoring method and system based on a physical information neural network, which comprises the following steps: collecting air quality data collected by an edge sensing device installed on the wall area of a target high and large space, collecting entrance airflow parameters collected by an airflow parameter measuring device installed at the entrance, and collecting door opening and closing state information; constructing a physical information neural network model, taking the pollutant release intensity as a learnable parameter, and adopting a dynamic diffusion coefficient; inputting the edge sensing data into the model as a boundary constraint, dynamically switching the boundary conditions according to the door opening and closing state, and inversely predicting the pollutant estimated concentration of the target high and large space; comparing the pollutant estimated concentration with a preset concentration threshold value, and generating an air quality alarm signal if the threshold value is exceeded. The present application can realize accurate monitoring and timely early warning of the air quality of a personnel activity area without adding a central suspended sensor.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and environmental monitoring technology, and in particular to a method and system for monitoring air quality in large spaces based on physical information neural networks. Background Technology

[0002] Highway service areas, high-speed rail waiting halls, and large commercial complexes are generally characterized by high single-story ceiling heights (typically 6 to 15 meters), large horizontal spans, and high pedestrian traffic. The main air pollutants in these tall spaces originate from carbon dioxide produced by human respiration and metabolic waste products emitted by the human body. To ensure indoor air quality, real-time monitoring of air quality in these spaces is necessary, along with timely warnings when air quality deteriorates.

[0003] In existing air quality monitoring technologies, environmental monitoring sensors (such as carbon dioxide sensors and temperature sensors) are typically limited by wiring costs, maintenance difficulties, and architectural aesthetic requirements, and can usually only be installed on walls or columns at the perimeter of a space, about 1.5 meters above the ground. However, human activity and pollutant release are mainly concentrated in the central area of ​​the space (such as rest areas and dining areas), which leads to a significant spatial misalignment between the sensor measurement location and the actual pollution source location. Due to the time lag in the diffusion and transport of pollutants in non-uniform flow fields, when the air quality in the central area begins to deteriorate, the readings of edge sensors often have not yet increased significantly. Conventional monitoring methods based on edge measurement data cannot accurately reflect the true air quality in areas of human activity, resulting in delayed or missed warnings.

[0004] Furthermore, the doors of such public buildings are often opened and closed frequently or left open for extended periods. When strong natural winds are present outdoors, wind pressure may force air into the building through the doors, creating airflow resistance in the opposite direction to the exhaust, leading to a significant reduction in the effective exhaust volume of the rooftop exhaust fans, or even exhaust failure. However, conventional sensors placed indoors cannot detect these dynamic changes in boundary conditions, and monitoring models therefore cannot accurately predict pollutant concentrations in the central area.

[0005] Some existing technologies attempt to use neural networks to predict indoor environmental parameters. For example, some solutions use neural networks to build predictive models for temperature or concentration, but these purely data-driven models usually require a large amount of historical data for training and are difficult to generalize to building spaces with different structures. Other solutions use computational fluid dynamics (CFD) simulations to analyze indoor flow field distribution, but CFD calculations are time-consuming and cannot meet the needs of real-time monitoring.

[0006] Therefore, there is an urgent need for a monitoring technology solution that can use edge measurement data to calculate the air quality in the central area in real time and issue early warnings without adding a suspended sensor in the center of the space. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for monitoring air quality in large spaces based on physical information neural networks, so as to solve the problems of edge sensors being unable to accurately reflect the air quality in the central area and the lag in monitoring and early warning in the prior art.

[0008] To achieve the above-mentioned objectives, the technical solution provided by this invention includes:

[0009] A high-altitude air quality monitoring method based on physical information neural networks includes the following steps:

[0010] Collect edge sensing data, including: air quality data collected by edge sensing devices installed on the walls of the target tall space, entrance airflow parameters and door opening / closing status information collected by airflow parameter measurement devices installed at the entrance of the target tall space;

[0011] A physical information neural network model is constructed to use pollutant release intensity as a learnable parameter and adopt a dynamic diffusion coefficient related to flow field velocity. By minimizing the weighted combination of data loss and physical equation residual loss, the estimated concentration of pollutants in the target high space is output.

[0012] The edge sensing data is used as boundary constraints and input into the physical information neural network model to inversely predict the estimated concentration of pollutants in the target tall space; wherein, the boundary conditions of the physical information neural network model are dynamically switched according to the opening and closing state of the gate.

[0013] Based on the difference between the estimated concentration of the pollutants and the preset target concentration, a feedback control algorithm is used to generate a frequency control command for the exhaust fan to adjust the exhaust volume of the exhaust system.

[0014] Preferably, the dynamic diffusion coefficient The calculation is based on the local airflow velocity field modulus and the inlet airflow parameter data, and the calculation formula is as follows:

[0015] ;

[0016] in, As a reference value for molecular diffusion, The velocity modulus of the flow field output by the physical information neural network model. For inlet wind speed, and This is an empirical coefficient.

[0017] Preferably, the boundary conditions for dynamically switching the physical information neural network model specifically include:

[0018] Set a gate state variable. When the gate state variable indicates that the gate is open, set the entrance boundary as a speed entrance boundary condition and use the entrance wind speed and wind direction as boundary parameters.

[0019] When the gate state variable indicates that the gate is closed, the entrance boundary is set to a no-slip wall boundary condition.

[0020] Preferably, the step of generating the exhaust fan frequency control command further includes feedforward air pressure compensation:

[0021] A wind pressure disturbance compensation lookup table is pre-constructed based on computational fluid dynamics (CFD) simulation to establish a mapping relationship between external natural wind pressure and the required compensation frequency;

[0022] When the inlet wind speed is detected to exceed the preset threshold, the corresponding exhaust attenuation amount is queried according to the wind pressure interference compensation lookup table, and the exhaust attenuation amount is converted into a frequency compensation amount and added to the frequency adjustment amount of the exhaust fan inverter.

[0023] Preferably, the step of generating the exhaust fan frequency control command further includes thermal buoyancy energy-saving correction:

[0024] Calculate the thermal pressure driving force generated by the temperature difference between indoors and outdoors in the target high-ceiling space, and quantify its natural contribution to the exhaust volume.

[0025] The energy-saving correction frequency deduction value is calculated based on the aforementioned natural contribution.

[0026] The energy-saving correction frequency deduction value is subtracted from the exhaust fan frequency control command to reduce fan energy consumption by utilizing the chimney effect.

[0027] Preferably, the method further includes a nighttime reference self-calibration step:

[0028] During the preset low-traffic period at night, the edge sensing data is collected as steady-state data to perform benchmark calibration on the physical information neural network model;

[0029] Based on the difference between the model's predicted value and the air quality data, the output bias parameters of the physical information neural network model are updated to eliminate drift errors caused by long-term operation.

[0030] Preferably, the physical information neural network model is deployed using transfer learning:

[0031] Construct a pre-trained base model that includes a physical constraint layer and a boundary adaptation layer that encode physical equations.

[0032] The network weights of the physical constraint layer are frozen, and the weights of the boundary adaptation layer are fine-tuned using a small number of samples collected on-site.

[0033] Preferably, the training process of the physical information neural network model employs a composite loss function, the composite loss function L of which is in the form of:

[0034] ;

[0035] in, The observation data loss term for edge sensing data. The physical residual loss term describes the convection-diffusion equation and is used to constrain the output to conform to the convection-diffusion equation. For boundary condition loss terms, and These are the weighting coefficients.

[0036] This invention also discloses a high-altitude air quality monitoring system based on a physical information neural network, comprising:

[0037] The sensing module includes an edge sensing device installed on the wall area of ​​the target tall space and an airflow parameter measuring device installed at the entrance of the target tall space. It is configured to collect edge sensing data, including air quality data, entrance airflow parameters and door opening and closing status information.

[0038] The calculation module is equipped with a physical information neural network model and is configured to receive the edge sensing data, use the physical information neural network model to invert and predict the estimated concentration of pollutants in the target tall space, and generate exhaust fan frequency control commands based on the difference between the estimated concentration of pollutants and the preset target concentration; wherein, the calculation module is also configured to dynamically switch the boundary conditions of the physical information neural network model based on the door opening and closing status information.

[0039] The execution module includes an exhaust fan frequency converter, configured to adjust the exhaust volume of the exhaust system in response to the exhaust fan frequency control command.

[0040] Beneficial effects

[0041] 1. This invention employs a physical information neural network model, using air quality data collected by edge sensing devices and inlet airflow parameter data as boundary constraints to inversely predict pollutant concentrations in high-ceilinged target spaces. This achieves accurate monitoring of air quality in areas with high population density without the need for additional central suspended sensors. Compared to purely data-driven prediction models, the physical information neural network embeds physical laws such as convection-diffusion equations into the model training process. This ensures that the model output strictly adheres to fluid dynamics principles while satisfying the observed data, significantly improving the physical rationality and generalization ability of the predictions.

[0042] 2. This invention solves for pollutant release intensity as a learnable parameter, enabling adaptive perception of the impact of population density changes on air quality without the need for population statistics. At the same time, by adopting a dynamic diffusion coefficient related to flow field velocity, the model can adapt to changes in turbulent transport characteristics under different operating conditions, thus improving the model's adaptability to dynamic environments.

[0043] 3. This invention dynamically switches the boundary conditions of the physical information neural network model based on the door's opening and closing status information, accurately capturing the impact of door opening and closing status changes on the indoor flow field, making the monitoring results more accurate and reliable. When the estimated concentration of pollutants exceeds a preset threshold, the system promptly generates an alarm signal, providing a basis for subsequent ventilation, personnel evacuation, and other response measures.

[0044] 4. In some preferred embodiments, this application also provides an exhaust linkage control function that responds to alarm signals. Combined with feedforward wind pressure compensation and thermal buoyancy energy-saving correction mechanism, it can effectively cope with external wind pressure interference and make reasonable use of natural thermal pressure, thus realizing the integration of monitoring and control. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a preferred embodiment of the high-altitude air quality monitoring method based on a physical information neural network according to the present invention.

[0046] Figure 2 This is a schematic diagram of the structure of a high-altitude air quality monitoring system based on a physical information neural network in a preferred embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0048] Example 1

[0049] The air quality monitoring method for tall spaces based on physical information neural networks provided in this embodiment can be implemented in an air quality monitoring system that includes an edge computing device, a main controller, and an alarm device. This system is suitable for tall public building spaces with a net height ranging from 6 to 15 meters, such as highway service areas, high-speed rail waiting halls, and large commercial complexes.

[0050] like Figure 1 As shown, the high-altitude air quality monitoring method based on physical information neural networks provided by this invention includes the following steps:

[0051] S1. Collect edge sensing data, including: air quality data collected by edge sensing devices installed on the walls of the target tall space, entrance airflow parameters and door opening / closing status information collected by airflow parameter measurement devices installed at the entrance of the target tall space.

[0052] In this step, the system periodically (e.g., every 30 seconds) collects real-time data from each sensor. Specifically:

[0053] Air quality data acquisition: Carbon dioxide sensors in the edge sensing device measure CO2 concentration in the wall area, while temperature sensors measure the air temperature at the corresponding location. Measurements from multiple sensors can be fused using weighted averaging or spatial interpolation to obtain representative values ​​for the edge area. Because the sensors are installed at the edge of the space rather than in the center, these measurements reflect the air conditions in the edge area rather than the area where people are active.

[0054] Inlet airflow parameter acquisition: The anemometer in the airflow parameter measurement device measures the wind speed and direction at the inlet. The wind speed measurement range is typically 0-20 m / s, and the wind direction is expressed as an angle relative to the inlet normal. These parameters are used to characterize the degree of influence of external natural wind on the indoor flow field.

[0055] Access control for the gate's open / closed status: A magnetic sensor detects the gate's open / closed status and outputs a binary signal: a 1 indicates the gate is open, and a 0 indicates the gate is closed. This status information is used to determine the type of entry boundary conditions in the physical information neural network model.

[0056] Through the aforementioned data acquisition, the system can obtain the real-time environmental status of the edge area of ​​a tall space, providing input for subsequent physical information neural network model inversion. Compared with the solution of setting up suspended sensors in the center of the space, the edge acquisition solution of this invention has advantages such as simple wiring, convenient maintenance, and no impact on the aesthetics of the building.

[0057] S2. Construct a physical information neural network model to use pollutant release intensity as a learnable parameter and adopt a dynamic diffusion coefficient related to flow field velocity. By minimizing the weighted combination of data loss and physical equation residual loss, the model outputs the estimated pollutant concentration in the target high space.

[0058] In existing technologies, the monitoring and prediction of indoor air quality in large spaces mainly fall into two categories:

[0059] The first category is purely data-driven machine learning methods. These methods use historically collected sensor data to train neural network models, establishing a mapping relationship between input (edge ​​measurements) and output (predicted values ​​for the central area). However, this method has the following drawbacks: it requires a large amount of historical data for training, resulting in long data collection cycles and high costs; the model lacks physical constraints, which may lead to abnormal outputs that violate fluid dynamics; and it has poor generalization ability for building spaces with different structures, requiring data to be collected and the model trained again for each space.

[0060] The second category is simulation methods based on computational fluid dynamics (CFD). These methods accurately calculate indoor flow and concentration field distributions by solving the Navier-Stokes equations and convection-diffusion equations. However, CFD calculations have the following drawbacks: they are computationally time-consuming, making it difficult to meet the needs of real-time monitoring; they are sensitive to boundary conditions and source term parameters, requiring these parameters to be accurately set in advance; and they have high computational resource requirements, making them difficult to deploy on edge computing devices.

[0061] To address the aforementioned shortcomings, this invention employs the Physics-Informed Neural Network (PINN) method, embedding physical equation constraints into the neural network training process. Unlike traditional purely data-driven models, PINN simultaneously considers the fitting error of the observed data and the residuals of physical laws in its optimization objective, ensuring that the model output conforms to both the measured data and physical laws.

[0062] The PINN model design used in this invention is described below:

[0063] 1. Model architecture design:

[0064] In this invention, the physical information neural network model adopts a fully connected neural network architecture. The network input includes spatiotemporal coordinates (x, y, z, t) and environmental parameter vector E, and the network output includes flow field velocity vector u=(u, v, w) and pollutant concentration field C.

[0065] In one specific embodiment, the neural network comprises six hidden layers, each containing 128 neurons, using the hyperbolic tangent (tanh) activation function. The input layer receives an 8-dimensional input vector (three-dimensional spatial coordinates, time, multiple edge concentration measurements, and inlet wind speed), and the output layer produces a 4-dimensional output vector (three-dimensional velocity components and concentration values).

[0066] Besides fully connected neural networks, physical information neural network models can also employ architectures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), or graph neural networks (GNNs). Activation functions, in addition to the hyperbolic tangent function, can include ReLU, Swish, or GELU. The number of layers and nodes in the network can be adjusted according to the complexity of the specific application scenario.

[0067] 2. Loss function design:

[0068] The training process of the physical information neural network model employs a composite loss function, the form of which is:

[0069] ;

[0070] in: This is the observation data loss term for edge sensing data, which measures the deviation of the network output from the measured value at the sensor location. To describe the physical residual loss term of the convection-diffusion equation, the constraint network output conforms to the fluid dynamics equations throughout the computational domain; The boundary condition loss term constrains the network output to satisfy the physical boundary conditions at the boundary. and These are weighting coefficients used to balance the relative importance of different loss terms.

[0071] In one specific embodiment, The value range is from 0.1 to 10. The value ranges from 0.01 to 1, and the specific value can be determined through hyperparameter search.

[0072] 3. Physical equation constraints:

[0073] The physical residual loss item This is used to constrain the output to conform to the convection-diffusion equation. Specifically, the convection-diffusion equation describes the transport process of pollutants in the flow field:

[0074] ;

[0075] Where C is the pollutant concentration, u is the flow field velocity vector, D is the diffusion coefficient, and S is the pollutant source term.

[0076] Physical residual loss term The calculation is performed using automatic differentiation techniques. Specifically, the automatic differentiation function built into the deep learning framework is used to calculate the partial derivatives of the neural network output C with respect to time t and spatial coordinates (x, y, z). These partial derivatives are then substituted into the convection-diffusion equation to obtain the equation residuals. The physical residual loss is the mean square value of the residuals.

[0077] ;

[0078] in, To calculate the number of collocation points within the domain, the collocation points are randomly sampled in the spatiotemporal domain or distributed according to a regular grid.

[0079] 4. Pollutant source terms as learnable parameters:

[0080] In existing technologies, the pollutant source term S(x,y,t) typically needs to be predetermined through personnel counting or other means, serving as known input parameters for the model. However, in real-world, large spaces, personnel distribution is dynamically changing and difficult to measure accurately in real time. Inaccurate source term setting can lead to significant deviations between the model output and the actual situation.

[0081] To address this problem, a key innovation of this invention lies in solving for the pollutant release intensity S as a learnable parameter (latent variable), rather than setting it as a fixed value or a pre-measured known quantity. This invention sets S as a trainable parameter of a neural network, uses measurement data from edge sensors as anchor points, and back-calculates the current pollutant source term intensity distribution by minimizing the physical residual.

[0082] Specifically, the source term S(x,y,t) can be represented as the output of another small neural network, or as a weighted combination of several basis functions:

[0083] ;

[0084] in, For predefined basis functions (such as Gaussian functions). These are the weight coefficients to be optimized. During training, the weights... The parameters of the main network are optimized together, so that the model can automatically infer a reasonable source term distribution while satisfying the edge measurement data.

[0085] In this way, the present invention can adaptively sense the impact of changes in crowd density on air quality without counting the number of people, thus achieving the effect of sensing crowd density without counting heads.

[0086] 5. Dynamic diffusion coefficient:

[0087] In existing technologies, the diffusion coefficient D in the convection-diffusion equation is usually set as a constant, taking the value of the molecular diffusion coefficient (approximately 1.6 x 10⁻⁶ for CO₂ diffusion in air). -5 m 2 / s). However, in high-ceilinged spaces, due to the operation of exhaust fans and the effect of external wind pressure, indoor airflow exhibits varying degrees of turbulence. Turbulence significantly enhances the transport and mixing of pollutants, with its effective diffusion coefficient being much larger than the molecular diffusion coefficient. Using a constant molecular diffusion coefficient underestimates the actual propagation speed of pollutants, leading to biased model predictions.

[0088] To address this issue, another key innovation of this invention lies in employing a dynamic diffusion coefficient that is related to the flow field velocity, rather than using a constant molecular diffusion coefficient. By setting the diffusion coefficient D as a function of the flow field velocity, this invention enables the model to adapt to the turbulent transport characteristics under different operating conditions.

[0089] In a preferred embodiment, the dynamic diffusion coefficient is calculated based on the local airflow velocity field modulus and the inlet airflow parameter data, and the calculation formula is as follows:

[0090] ;

[0091] in, This is the baseline value for molecular diffusion; for CO2 diffusion in air, the typical value is approximately 1.6 x 10⁻⁵ m² / s. The velocity modulus of the flow field output by the physical information neural network model; Inlet wind speed; and This is an empirical coefficient, which can be calibrated using experimental or simulation data.

[0092] In one specific embodiment, The value range is from 0.01 to 0.1. The value ranges from 0.001 to 0.01. The physical meaning of this formula is that when the local velocity or inlet wind speed increases, the turbulence intensity increases, and the effective diffusion coefficient increases accordingly, thus more accurately describing the rapid mixing process of pollutants.

[0093] By employing a dynamic diffusion coefficient, this invention improves the model's adaptability to dynamic environments, enabling the model to provide accurate concentration predictions under different exhaust fan speeds and external wind conditions.

[0094] In summary, this invention, by employing a physical information neural network model, embeds physical laws such as convection-diffusion equations into the model training process, compared to purely data-driven prediction models. This ensures that the model output strictly adheres to fluid dynamics laws while satisfying observational data, significantly improving the physical rationality and generalization ability of the predictions. The model exhibits stronger generalization ability and better physical interpretability, avoiding the abnormal outputs that violate physical laws that may occur with purely data-driven models.

[0095] S3. Using the edge sensing data as boundary constraints, input the physical information neural network model to invert and predict the estimated concentration of pollutants in the target tall space; wherein, the boundary conditions of the physical information neural network model are dynamically switched according to the opening and closing state of the gate.

[0096] In this step, real-time edge sensing data is input into a pre-trained physical information neural network model. The model uses edge measurement data as boundary constraints and, guided by physical equations, inversely calculates the pollutant concentration distribution throughout the space (especially the central area where people are active).

[0097] It should be noted that the air quality data collected by the edge sensing device is used as the boundary constraint condition for the physical information neural network model. Specifically, during model inference, the measured values ​​at the sensor locations are taken as known conditions, requiring the model output to be consistent with the measured values ​​at these locations. This is equivalent to providing anchor points for the inversion problem, enabling the model to infer the concentration distribution throughout the space based on limited edge observation data.

[0098] In existing technologies, indoor environment models typically employ fixed boundary condition settings. For example, door and window positions are uniformly set as open or closed boundaries. However, in real-world high-ceilinged spaces, the opening and closing states of doors are dynamic and significantly impact the indoor airflow. When a door is closed, the entrance area acts like a solid wall, preventing air penetration; when the door is open, the entrance becomes an open channel, allowing outside air to flow into the room under wind pressure. Using fixed boundary conditions fails to accurately describe this dynamic change, leading to increased prediction bias when door states change.

[0099] To address this issue, in a preferred embodiment of the present invention, the dynamic switching of the boundary conditions of the physical information neural network model specifically includes: setting a gate state variable; when the gate state variable indicates that the gate is open, setting the entrance boundary as a velocity entrance boundary condition, and using the entrance wind speed and wind direction as boundary parameters; when the gate state variable indicates that the gate is closed, setting the entrance boundary as a non-slip wall boundary condition.

[0100] This invention introduces the gate state variable K door Achieve dynamic switching of boundary conditions: when K door When the velocity is 0 (the gate is closed), in the physical information neural network model, the entrance boundary adopts a no-slip wall boundary condition, meaning the velocity at the boundary is zero. When K door When =1 (door open), the inlet boundary condition switches to velocity inlet boundary condition, and the wind speed v measured by the airflow parameter measuring device is used. in Assign the wind direction θ to this boundary: .

[0101] Through this dynamic boundary condition switching mechanism, the present invention enables the physical information neural network model to accurately capture the impact of changes in the opening and closing state of the door on the indoor flow field, thereby improving the accuracy of monitoring.

[0102] After the model inference is completed, the estimated concentration of pollutants in the target high space is output. In particular, the system extracts the concentration prediction value of the central personnel activity area (e.g., the horizontal center area at a height of 1.5 meters above the ground) as input for alarm judgment and optional control algorithms.

[0103] Through the aforementioned inversion prediction, this invention achieves accurate monitoring of air quality in areas with high population density without the need for additional central suspended sensors. It solves the problem of sensor misalignment with the spatial distribution of activity areas. Compared to traditional methods that rely solely on edge data for monitoring, this invention can more accurately perceive the actual air quality in these areas, making monitoring and early warning more timely and precise.

[0104] Step S4: Compare the estimated concentration of the pollutant with a preset concentration threshold. If the concentration exceeds the preset concentration threshold, generate an air quality alarm signal.

[0105] In this step, the estimated pollutant concentration output by the physical information neural network model is compared with the preset concentration threshold to determine whether an alarm needs to be generated.

[0106] 1. Comparison of concentration thresholds:

[0107] Estimating pollutant concentrations With preset concentration threshold Comparison:

[0108] Alarm conditions: ;

[0109] In one specific embodiment, the concentration threshold The recommended CO2 concentration is set at 1000 ppm, which is the warning level recommended in indoor air quality standards. When the estimated concentration exceeds this threshold, it indicates that the air quality in the central activity area has deteriorated and measures need to be taken.

[0110] 2. Alarm signal generation:

[0111] When detected When this occurs, an air quality alarm signal is generated. The alarm signal can trigger one or more of the following response methods:

[0112] The audible and visual alarm sounds an alert.

[0113] An alarm window popped up on the monitoring center's display screen;

[0114] Send notifications to administrators via SMS or app;

[0115] The exhaust control system is activated in conjunction with the system (see step S5).

[0116] Trigger the personnel evacuation plan (applicable to situations where the level is severely exceeded).

[0117] 3. Multi-level alarm mechanism:

[0118] In a preferred embodiment, a multi-level alarm mechanism can be employed:

[0119] Level 1 warning ( ): Only logs and push notifications are recorded;

[0120] Level 2 Alert ( ): Activate the audible and visual alarm and link it to the exhaust system;

[0121] Level 3 Alert ): Initiate emergency ventilation and notify relevant management departments.

[0122] By using alarms to assess air quality, this application enables timely early warning of air quality in large spaces, providing a basis for subsequent response measures.

[0123] In some preferred embodiments, in response to the air quality alarm signal, the method further includes an exhaust ventilation linkage control step:

[0124] S5. Based on the difference between the estimated concentration of the pollutants and the preset target concentration, a frequency control command for the exhaust fan is generated through a feedback control algorithm to adjust the exhaust volume of the exhaust system.

[0125] In existing technologies, exhaust control systems typically rely solely on feedback control, adjusting the exhaust fan output based on concentration deviations measured by sensors. However, when doors are open and strong natural winds are present outdoors, external wind pressure may oppose the exhaust fan's suction direction. If the external wind pressure is sufficiently high, it can cause a significant decrease in the effective exhaust volume of the fan, or even lead to air backflow. This disturbance is sudden and time-varying, while feedback control requires waiting for the disturbance to manifest in concentration changes before responding, exhibiting a significant lag.

[0126] In this embodiment, the estimated pollutant concentration output by the physical information neural network model is compared with the preset target concentration to calculate the concentration deviation, and a frequency control command for the exhaust fan is generated through a feedback control algorithm.

[0127] When the deviation is greater than 0, it means that the current air quality is lower than the target requirement and the exhaust volume needs to be increased; when the deviation is less than 0, it means that the air quality is better than the target and the exhaust volume can be appropriately reduced to save energy.

[0128] The basic demand frequency is calculated using feedback control algorithms (such as PID control). :

[0129] ;

[0130] in, , , These are the proportional, integral, and derivative coefficients, which can be tuned according to the system characteristics.

[0131] In addition to PID control algorithms, feedback control can also employ methods such as fuzzy control, model predictive control (MPC), or reinforcement learning control.

[0132] In existing technologies, exhaust control systems typically rely solely on feedback control, adjusting the exhaust fan output based on concentration deviations measured by sensors. However, when doors are open and strong natural winds are present outdoors, external wind pressure may oppose the exhaust fan's suction direction. If the external wind pressure is sufficiently high, it can cause a significant decrease in the effective exhaust volume of the fan, or even lead to air backflow. This disturbance is sudden and time-varying, while feedback control requires waiting for the disturbance to manifest in concentration changes before responding, exhibiting a significant lag.

[0133] To address this issue, in a preferred embodiment, the step of generating the exhaust fan frequency control command further includes feedforward wind pressure compensation: a wind pressure interference compensation lookup table is pre-constructed based on computational fluid dynamics (CFD) simulation to establish a mapping relationship between external natural wind pressure and the required compensation frequency; when the inlet wind speed is detected to exceed a preset threshold, the corresponding exhaust attenuation is queried according to the wind pressure interference compensation lookup table, and the exhaust attenuation is converted into a frequency compensation amount and superimposed on the frequency adjustment amount of the exhaust fan inverter.

[0134] Specifically, before system deployment, computational fluid dynamics (CFD) software was used to simulate the target tall space, calculating the effective exhaust volume attenuation rate of the exhaust fans under different combinations of inlet wind speed and direction. The simulation results formed a two-dimensional lookup table, with the input to the table being... The output is the exhaust attenuation rate. For example, a lookup table entry might be: when the inlet wind speed... ,wind direction When facing the entrance (i.e., directly opposite), the exhaust attenuation rate is 15%, and the corresponding compensation frequency increment is... .

[0135] During system operation, when the airflow parameter measuring device detects that the inlet wind speed exceeds a preset threshold (e.g., 2 m / s), the system, based on real-time data... The corresponding frequency compensation amount can be obtained by querying the preset wind pressure interference compensation lookup table. And then add it to the basic demand frequency.

[0136] In addition to lookup table methods, feedforward compensation can also be calculated using simplified physical formulas or by using another feedforward neural network for prediction.

[0137] Through this feedforward compensation mechanism, the present invention can respond instantly to wind pressure interference, without waiting for concentration changes to be reflected in sensor readings, significantly improving the system's anti-interference capability. This allows the exhaust system to maintain stable ventilation even under conditions of frequent door opening and closing or changes in outdoor wind pressure.

[0138] In existing technologies, exhaust control systems typically output the full mechanical exhaust power according to the set exhaust demand, without considering the contribution of natural ventilation. However, in high-ceilinged spaces, when the indoor air temperature is higher than the outdoor temperature, the density of hot air is less than that of cold air, generating upward buoyancy. This thermo-pressure effect (also known as the chimney effect) naturally drives the hot air upward and out through the top, effectively providing additional auxiliary power for mechanical exhaust. If the contribution of thermo-pressure is not considered, exhaust fans will still operate at high power when the thermo-pressure is strong, resulting in unnecessary energy waste.

[0139] To address this issue, in a preferred embodiment, the step of generating the exhaust fan frequency control command further includes thermal buoyancy energy-saving correction: calculating the thermal pressure driving force generated by the indoor-outdoor temperature difference within the target tall space, quantifying its natural contribution to the exhaust volume; converting the natural contribution into an energy-saving correction frequency deduction value; subtracting the energy-saving correction frequency deduction value from the exhaust fan frequency control command, thereby reducing fan energy consumption using the chimney effect.

[0140] Based on the physical principle of the chimney effect, thermal pressure driving force It can be represented as:

[0141] ;

[0142] in, air density, It is the acceleration due to gravity. For the clear height of the space, Due to the temperature difference between indoors and outdoors, The mean absolute temperature.

[0143] After converting the thermal pressure driving force into an equivalent exhaust volume contribution, the corresponding energy-saving correction frequency deduction value can be obtained. .

[0144] By deducting the natural contribution of thermal pressure, this invention enables the system to reduce the mechanical power output of the exhaust fan while ensuring effective ventilation, thus achieving energy-saving operation. This energy-saving effect is particularly significant during seasons with large indoor-outdoor temperature differences (such as the winter heating season). Combined with on-demand ventilation control based on concentration feedback, this invention achieves significant energy savings compared to traditional fixed-frequency or simple threshold control schemes.

[0145] By integrating feedback control, feedforward air pressure compensation, and thermal buoyancy energy-saving correction, the final output exhaust fan frequency control command is as follows:

[0146] ;

[0147] Frequency control command The command is sent to the exhaust fan's frequency converter, which adjusts the fan's operating frequency accordingly, thereby changing the exhaust volume. The effective range of the frequency control command is typically 0-50Hz, corresponding to continuous adjustment of the exhaust fan from stop to full-load operation.

[0148] Furthermore, this invention also includes a nighttime reference self-calibration step. Specifically, in existing technologies, sensors may develop drift errors during long-term operation, causing measured values ​​to gradually deviate from the true values. In addition, the environmental model itself may gradually develop systematic deviations due to slow changes in environmental conditions (such as building settlement, changes in furniture layout, etc.). These accumulated errors lead to a gradual decrease in monitoring accuracy, requiring periodic manual calibration.

[0149] To address this issue, in a preferred embodiment, the method further includes a nighttime baseline self-calibration step: during a preset nighttime low-traffic period, the edge sensing data is collected in real time as steady-state data to perform baseline calibration on the physical information neural network model; based on the difference between the model's predicted value and the air quality data, the output bias parameters of the physical information neural network model are updated to eliminate drift errors generated during long-term operation.

[0150] This invention selects a low-traffic period at night (e.g., 3:00-4:00 AM) for self-calibration. At this time, there is almost no human activity in the space, pollutant source terms are close to zero, and the indoor air condition is close to a steady state of equilibrium with the outdoor air. Under these conditions, the concentrations in the edge and central areas should tend to be consistent and close to the background concentration of outdoor air (approximately 400 ppm).

[0151] During the self-calibration period, the system collects edge sensing data as a steady-state reference and simultaneously runs a physical information neural network model to obtain predicted values ​​for the central region. Since it is assumed that the actual concentration distribution is uniform at this time, the difference between the model prediction and the edge measurement can be attributed to systematic bias in the model.

[0152] The system calculates the difference between the model's predicted value and the measured value, and updates the model's output bias parameter accordingly, so that the calibrated model outputs a result consistent with the measured value under steady-state conditions.

[0153] Through this periodic benchmark calibration mechanism, this application can effectively eliminate long-term cumulative errors caused by sensor drift and model deviation, ensuring long-term stable monitoring of the system and maintaining monitoring accuracy without manual intervention.

[0154] It should be understood that deploying machine learning models to new application scenarios typically requires collecting a large amount of new data and retraining the model from scratch in that scenario. For air quality monitoring applications in large spaces, different spaces have different geometric dimensions, inlet locations, and airflow organization patterns. If a physical information neural network model is trained from scratch for each space, it would require a large amount of field data and computing resources, resulting in long deployment cycles, high costs, and difficulty in large-scale deployment.

[0155] To address this issue, in a preferred embodiment, the physical information neural network model is deployed using transfer learning: a pre-trained basic model is constructed, which includes a physical constraint layer and a boundary adaptation layer that encode physical equation constraints; the network weights of the physical constraint layer are frozen, and the weights of the boundary adaptation layer are fine-tuned using a small number of samples collected on-site.

[0156] Specifically, a general-purpose, large-space PINN basic model is first trained in the cloud using a large amount of CFD simulation data. The network structure of this basic model is divided into two parts: a physical constraint layer, which encodes physical laws such as convection-diffusion equations, and these layers learn general physical laws that are independent of specific spaces; and a boundary adaptation layer, which learns the boundary characteristics and geometric features of a specific space.

[0157] When deploying the model to a specific target high space, the following steps are performed: First, input the basic geometric parameters of the target space (length, width, height, entrance location, etc.); then freeze the network weights of the physical constraint layer so that they remain unchanged during fine-tuning; next, use sensor data collected in the first 24-48 hours after on-site installation to fine-tune the weights of only the boundary adaptation layer; after fine-tuning, the model can be used for real-time inference in that space.

[0158] Through this transfer learning strategy, the present invention significantly reduces the data requirements and time costs of model deployment, enabling the rapid deployment of physical information neural network technology to various application scenarios. Furthermore, since the sensors only need to be installed at the edge of the space, eliminating the need for suspended supports and wiring in the central area, the engineering implementation cost is lower.

[0159] Example 2

[0160] like Figure 2 As shown, this invention also provides a high-altitude air quality monitoring system based on physical information neural networks, comprising three levels: a sensing module, a computing module, and an alarm device.

[0161] The perception module includes:

[0162] Edge sensing device: A sensor array installed on the walls of a large target space to collect air quality data. In one specific embodiment, the edge sensing device includes multiple carbon dioxide concentration sensors and temperature sensors, distributed and installed on the surrounding walls of the space at a height of approximately 1.5 meters above the ground. The number of sensors can be determined based on the space area, for example, one set of sensors per 200 square meters. In addition to carbon dioxide sensors, the edge sensing device can also employ other air quality sensors such as volatile organic compound (VOC) sensors and particulate matter (PM2.5 / PM10) sensors.

[0163] Airflow parameter measurement device: A measuring device installed at the entrance of the target tall space to collect entrance airflow parameters and door opening / closing status information. In one specific embodiment, the airflow parameter measurement device includes an ultrasonic anemometer and a door magnetic sensor, installed above or on both sides of the door frame at the main entrance of the building. Besides the ultrasonic anemometer, the airflow parameter measurement device can also use a hot-wire anemometer, a rotor anemometer, etc. Door status detection can also employ infrared beam sensors, video image analysis, etc., in addition to door magnetic sensors.

[0164] The calculation module includes:

[0165] Edge computing device: An industrial-grade computing gateway with an embedded AI acceleration chip, used to run physical information neural network models. In one specific embodiment, the edge computing device employs an embedded processor with neural network inference acceleration capabilities, supporting model deployment using deep learning frameworks such as PyTorch or TensorFlow Lite. Besides dedicated industrial gateways, edge computing devices can also utilize general-purpose embedded computers (such as Raspberry Pi, NVIDIA Jetson, etc.), industrial PCs, or cloud-distributed computing servers. When cloud computing is used, edge sensors upload data to the cloud via an IoT platform, and after completing model inference, the cloud sends control commands to the field controller.

[0166] Main controller: A programmable logic controller (PLC) or industrial computer used to perform alarm judgments and optional control algorithms. The main controller communicates with edge computing devices and alarm devices via industrial Ethernet or RS485 bus.

[0167] Alarm device: Configured to respond to air quality alarm signals and output alarm prompts. In one specific embodiment, the alarm device may include an audible and visual alarm, a display screen prompt, SMS / APP push notifications, etc. When the estimated concentration of pollutants is detected to exceed a preset threshold, the alarm device is triggered to remind relevant personnel to take measures.

[0168] In some preferred embodiments, an execution module is further included, the execution module comprising:

[0169] Exhaust fan frequency converter: A variable frequency drive device connected to the top exhaust fan, used to adjust the exhaust fan speed according to frequency control commands. In one specific embodiment, the frequency adjustment range of the frequency converter is 0-50Hz, corresponding to the continuous adjustment of the exhaust fan speed from stop to rated speed. In some preferred embodiments, after the system generates an alarm signal, it can also automatically adjust the exhaust volume in conjunction with the execution module.

[0170] It should be noted that the sensing module is configured to collect edge sensing data, including air quality data, inlet airflow parameters, and door opening / closing status information. The calculation module deploys a physical information neural network model, configured to receive the edge sensing data, use the physical information neural network model to inversely predict the estimated concentration of pollutants in the target tall space, and compare the estimated pollutant concentration with a preset concentration threshold. If the threshold is exceeded, an air quality alarm signal is generated. The calculation module is also configured to dynamically switch the boundary conditions of the physical information neural network model based on the door opening / closing status information. The alarm module is configured to respond to the air quality alarm signal and output an alarm prompt.

[0171] In some preferred embodiments, the system further includes an execution module, including an exhaust fan frequency converter, configured to adjust the exhaust volume of the exhaust system by generating an exhaust fan frequency control command by a calculation module in response to an air quality alarm signal.

[0172] In one specific embodiment, the edge sensing device includes 4 to 8 CO2 sensor nodes, distributed and installed on the surrounding walls of the space at a height of 1.5 meters above the ground. Each sensor node transmits measurement data to the edge computing device via wireless communication (such as LoRa or ZigBee). An airflow parameter measurement device is installed above the main entrance door frame, including an integrated wind speed and direction sensor and a door magnetic switch. The computing module consists of an edge computing device and a main controller (PLC). The edge computing device is responsible for running the inference process of the physical information neural network model, while the main controller is responsible for alarm judgment, optional PID control algorithms, feedforward compensation calculations, and frequency command output. The two communicate via industrial Ethernet.

[0173] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring air quality in large spaces based on physical information neural networks, characterized in that, Includes the following steps: Collect edge sensing data, including: air quality data collected by edge sensing devices installed on the walls of the target tall space, entrance airflow parameters and door opening / closing status information collected by airflow parameter measurement devices installed at the entrance of the target tall space; A physical information neural network model is constructed to use pollutant release intensity as a learnable parameter and adopt a dynamic diffusion coefficient related to flow field velocity. By minimizing the weighted combination of data loss and physical equation residual loss, the estimated concentration of pollutants in the target high space is output. The edge sensing data is used as boundary constraints and input into the physical information neural network model to inversely predict the estimated concentration of pollutants in the target tall space; wherein, the boundary conditions of the physical information neural network model are dynamically switched according to the opening and closing state of the gate. The estimated concentration of the pollutant is compared with a preset concentration threshold. If the concentration exceeds the preset concentration threshold, an air quality alarm signal is generated. The dynamic diffusion coefficient The calculation is based on the local airflow velocity field modulus and the inlet airflow parameter data, and the calculation formula is as follows: ； in, As a reference value for molecular diffusion, The velocity modulus of the flow field output by the physical information neural network model. For inlet wind speed, and This is an empirical coefficient; It also includes an exhaust linkage control step in response to the air quality alarm signal: Based on the difference between the estimated concentration of the pollutants and the preset target concentration, a feedback control algorithm is used to generate a frequency control command for the exhaust fan to adjust the exhaust volume of the exhaust system. The step of generating the exhaust fan frequency control command also includes feedforward air pressure compensation: A wind pressure disturbance compensation lookup table is pre-constructed based on computational fluid dynamics (CFD) simulation to establish a mapping relationship between external natural wind pressure and the required compensation frequency; When the inlet wind speed is detected to exceed the preset threshold, the corresponding exhaust attenuation amount is queried according to the wind pressure interference compensation lookup table, and the exhaust attenuation amount is converted into a frequency compensation amount and added to the frequency adjustment amount of the exhaust fan inverter.

2. The high-altitude air quality monitoring method based on physical information neural networks according to claim 1, characterized in that, The specific boundary conditions for dynamically switching the physical information neural network model include: Set a gate state variable. When the gate state variable indicates that the gate is open, set the entrance boundary as a speed entrance boundary condition and use the entrance wind speed and wind direction as boundary parameters. When the gate state variable indicates that the gate is closed, the entrance boundary is set to a no-slip wall boundary condition.

3. The high-altitude air quality monitoring method based on physical information neural networks according to claim 1, characterized in that, The step of generating the exhaust fan frequency control command also includes thermal buoyancy energy-saving correction: Calculate the thermal pressure driving force generated by the temperature difference between indoors and outdoors in the target high-ceiling space, and quantify its natural contribution to the exhaust volume. The energy-saving correction frequency deduction value is calculated based on the aforementioned natural contribution. The energy-saving correction frequency deduction value is subtracted from the exhaust fan frequency control command to reduce fan energy consumption by utilizing the chimney effect.

4. The high-altitude air quality monitoring method based on physical information neural networks according to claim 1, characterized in that, The method also includes a nighttime reference self-calibration step: During the preset low-traffic period at night, the edge sensing data is collected as steady-state data to perform benchmark calibration on the physical information neural network model; Based on the difference between the model's predicted value and the air quality data, the output bias parameters of the physical information neural network model are updated to eliminate drift errors caused by long-term operation.

5. The method for monitoring air quality in large spaces based on a physical information neural network according to claim 1, characterized in that, The physical information neural network model is deployed using transfer learning: Construct a pre-trained base model that includes a physical constraint layer and a boundary adaptation layer that encode physical equations. The network weights of the physical constraint layer are frozen, and the weights of the boundary adaptation layer are fine-tuned using a small number of samples collected on-site.

6. The method for monitoring air quality in large spaces based on physical information neural networks according to claim 1, characterized in that, The training process of the physical information neural network model employs a composite loss function, the form of which is: ； in, The observation data loss term for edge sensing data. The physical residual loss term describes the convection-diffusion equation and is used to constrain the output to conform to the convection-diffusion equation. For boundary condition loss terms, and These are the weighting coefficients.

7. A system for implementing the high-altitude air quality monitoring method based on a physical information neural network as described in any one of claims 1-6, characterized in that, include: The sensing module includes an edge sensing device installed on the wall area of ​​the target tall space and an airflow parameter measuring device installed at the entrance of the target tall space. It is configured to collect edge sensing data, including air quality data, entrance airflow parameters and door opening and closing status information. The calculation module is equipped with a physical information neural network model and is configured to receive the edge sensing data, use the physical information neural network model to invert and predict the estimated concentration of pollutants in the target tall space, and generate exhaust fan frequency control commands based on the difference between the estimated concentration of pollutants and the preset target concentration; wherein, the calculation module is also configured to dynamically switch the boundary conditions of the physical information neural network model based on the door opening and closing status information. The alarm module is configured to output an alarm prompt in response to the air quality alarm signal.

Images