Spray distribution prediction method and system based on adaptive polynomial random forest

Abstract

The invention relates to the technical field of environmental health monitoring, and discloses a droplet distribution prediction method and system based on an adaptive polynomial random forest, and the method comprises the steps: firstly simulating the droplet distribution condition based on a set simulation condition, and constructing a droplet distribution prediction model according to a simulation result; the droplet distribution is predicted based on the constructed droplet distribution prediction model according to the real-time cough time, direction and size, a numerical calculation method is combined with an ML algorithm, and compared with a traditional numerical simulation method, the method can greatly save calculation time and cost and has excellent prediction precision.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental sanitation monitoring, and in particular to a droplet distribution prediction method and system based on adaptive polynomial random forest. Background Art

[0002] Droplet transmission is one of the main transmission routes for many respiratory diseases. Accurately predicting the distribution of droplets in the air is crucial for assessing the risk of disease transmission and formulating prevention and control measures. At present, domestic and foreign scholars generally use experimental and numerical simulation methods to study cough droplet transmission. Due to the complexity of processes such as droplet movement and evaporation, it can be challenging to control experimental conditions and obtain data, resulting in limited existing data that cannot fully reflect reality. Traditional numerical simulation methods rely heavily on the experience and expertise of designers, often requiring a lot of time and resources to calculate, and are difficult to meet the needs of rapid response. Therefore, by using scientific methods to quickly predict the transmission pattern of cough droplets, especially in closed spaces or crowded areas, environmental control and air quality management in the field of public health can be strengthened, thereby accurately assessing the risk of virus transmission.

[0003] In order to address these limitations of traditional models, researchers introduced machine learning (ML) into particle distribution prediction in fluid dynamics, which has significant advantages over traditional methods in terms of computational efficiency. Although the above studies have conducted detailed research on particle diffusion, they are limited by environmental parameters and lack in-depth research on the spatiotemporal distribution patterns of droplets at the data level, resulting in incomplete reflection of the true mechanism of droplet transmission, slow response to the prediction of droplet transmission, and low prediction accuracy. Summary of the invention

[0004] The present invention provides a droplet distribution prediction method and system based on adaptive polynomial random forest to solve the problems in the prior art of incomplete reflection of the real mechanism of droplet transmission, slow prediction response of droplet transmission situation and low prediction accuracy.

[0005] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0006] In a first aspect, the present invention provides a method for predicting droplet distribution based on an adaptive polynomial random forest, comprising:

[0007] S1: setting droplet simulation conditions, and simulating droplet distribution based on the simulation conditions;

[0008] S2: constructing grid cells based on the simulation conditions, and independently calculating and analyzing the number and concentration of particles in each grid cell;

[0009] S3: performing two-dimensional Gaussian function fitting based on the number and concentration of the particles, and constructing N droplet distribution databases based on the fitting results, each of the droplet distribution databases including Gaussian fitting parameters, particle size, and time and space;

[0010] S4: Obtaining a target Gaussian model based on the N droplet distribution databases;

[0011] S5: Solving and verifying the target Gaussian model to obtain a final droplet distribution prediction model, and predicting droplet distribution according to the real-time cough time, direction and size based on the droplet distribution prediction model.

[0012] In a second aspect, the present application provides a droplet distribution prediction system based on an adaptive polynomial random forest, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect when executing the computer program.

[0013] Beneficial effects:

[0014] The droplet distribution prediction method based on adaptive polynomial random forest provided by the present invention first simulates the droplet distribution based on set simulation conditions, constructs a droplet distribution prediction model according to the simulation results, and predicts the droplet distribution according to the time, direction and size of the real-time cough based on the constructed droplet distribution prediction model. The Gaussian function model is combined with the random forest algorithm. Compared with the traditional numerical simulation method, it can greatly save calculation time and cost and has excellent prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of a droplet distribution prediction method based on adaptive polynomial random forest according to a preferred embodiment of the present invention;

[0016] Figure 2 This is a schematic diagram of the indoor 4×4×3m calculation domain model;

[0017] Figure 3 The spatiotemporal distribution diagram of cough droplets with different particle sizes ranging from 10 to 60 μm (5 μm interval);

[0018] Figure 4 The concentration of 10um cough droplet particles in the range of 0.4-0.6m at 6s;

[0019] Figure 5 The Gaussian function concentration fitting distribution diagram of 10um in the range of 0.4-0.6m at 6s;

[0020] Figure 6 This is the algorithm structure diagram of the spatiotemporal distribution of cough droplets of different particle sizes based on the Gaussian function model of APRF;

[0021] Figure 7 This is the concentration distribution diagram of 18um cough droplets in the x-axis direction of 0.8-1.0m for 6s;

[0022] Figure 8 This is a comparison chart of the predicted concentration and original concentration of 18um cough droplets of 6s cut along the y and z directions in the x-axis 0.8-1.0m direction. DETAILED DESCRIPTION

[0023] The technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the usual meanings understood by persons with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not indicate quantity restrictions, but indicate the existence of at least one. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship also changes accordingly.

[0025] See also Figure 1 , the present application provides a droplet distribution prediction method based on adaptive polynomial random forest, comprising:

[0026] S1: setting droplet simulation conditions, and simulating droplet distribution based on the simulation conditions;

[0027] S2: constructing grid cells based on the simulation conditions, and independently calculating and analyzing the number and concentration of particles in each grid cell;

[0028] S3: performing two-dimensional Gaussian function fitting based on the number and concentration of the particles, and constructing N droplet distribution databases based on the fitting results, each of the droplet distribution databases including Gaussian fitting parameters, particle size, and time and space;

[0029] S4: Obtaining a target Gaussian model based on the N droplet distribution databases;

[0030] S5: Solving and verifying the target Gaussian model to obtain a final droplet distribution prediction model, and predicting droplet distribution according to the real-time cough time, direction and size based on the droplet distribution prediction model.

[0031] The above-mentioned droplet distribution prediction method based on adaptive polynomial random forest first simulates the droplet distribution based on the set simulation conditions, constructs a droplet distribution prediction model according to the simulation results, and predicts the droplet distribution according to the time, direction and size of the real-time cough based on the constructed droplet distribution prediction model. The basic database of some Gaussian functions is provided by CFD, including the time position information of the particles, and the adaptive random forest algorithm is combined with it. By constructing a Gaussian function model of the existing data and using it as an algorithm learning sample for the machine to learn, it is possible to predict the distribution of different particle sizes in different time and space. Compared with the traditional numerical simulation method, it can greatly save computing time and cost, and has excellent prediction accuracy.

[0032] Optionally, the S1 includes:

[0033] In the fluid mechanics simulation software, set the preset spatial range;

[0034] The particle size range of the droplets simulated based on the Euler-Lagrange model within the preset spatial range is N1μm to N2μm, with an interval of nμm, and the distribution of cough droplets within a set time period under a set ambient temperature and a set relative humidity.

[0035] In this optional embodiment, the fluid dynamics simulation software can be Fluent, which is only used as an example and not limited here. Specifically, in one example, when defining the particle size range of cough droplets, N1 is 10 μm, N2 is 60 μm, the interval is 5 μm, the preset space range is 4×4×3m, the temperature and relative humidity are set to 25°C and 0% respectively, and the time period is set to a time period of 0-15s.

[0036] Optionally, the S2 includes:

[0037] Determine a computational domain based on the simulation conditions, and divide the computational domain into a plurality of grid units according to a preset division condition;

[0038] Count and analyze the number and concentration of particles in each grid cell;

[0039] The preset division conditions are as follows:

[0040] The numerical stability in each grid cell is higher than a set stability threshold, and the grid density is higher than a set density threshold, wherein the density threshold is determined based on whether subtle changes in droplets can be captured.

[0041] In this optional embodiment, a three-dimensional grid generation tool is used to divide the computational domain into 20×80×160 grid cells with a total volume of 4×4×3 meters. When determining the size of each grid, the numerical stability in each grid cell is taken as a benchmark to be higher than the set stability threshold, so as to achieve the purpose of ensuring numerical stability. It is worth pointing out that since the size of the grid density determines whether the subtle changes of the droplets can be captured, when determining the density of the grid, the grid density is higher than the set density threshold, so that the density of the divided grid is sufficient to capture the subtle changes of the droplets.

[0042] Specifically, the calculation and analysis of the number and concentration of particles in each grid unit includes:

[0043] In each time step, the number and concentration of cough droplets in each grid cell are recorded, and the concentration data at all times are summarized to form a time series data set of concentration distribution, and the concentration distribution of cough droplets in the X, Y, and Z directions is recorded.

[0044] In this simulation step, the particle size range of cough droplets is defined as 10μm to 60μm, with an interval of 5μm, and the initial velocity is set to 0.11m / s to simulate the actual release speed of droplets during coughing. In terms of time step configuration, the time range is set to 0 to 15 seconds, and the time step is 0.01 seconds to ensure that the time resolution is high enough to capture the dynamic behavior of droplets over time. In terms of simulation algorithm selection, the Euler-Lagrangian model is selected to track the movement of particles. This model will take into account the forces acting on particles in the flow field, including gravity, air resistance, etc. Further, the concentration distribution analysis is performed. First, the data collection process is to record the number and concentration of cough droplets in each grid in each time step. Ensure the accuracy and real-time nature of data collection for subsequent analysis. Then, recording the concentration distribution includes: summarizing the concentration data at each moment to form a time series data set, and recording the concentration distribution in the X, Y, and Z directions.

[0045] As a convertible implementation, in a feasible implementation, when the concentration is higher than a set value, for example, higher than 1% of the total number, the grid corresponding to the concentration can be further divided more finely. In this way, by introducing a finer grid division into the high-concentration area, dynamic division of the grid can be achieved, which can improve the calculation accuracy.

[0046] Furthermore, a two-dimensional Gaussian function is fitted to the concentration distribution of a space composed of an X-direction grid and the Y-axis and Z-axis with 1s as a moment, and N droplet distribution databases are obtained, which satisfy the following relationship:

[0047]

[0048] In the formula, a represents the amplitude, y represents the y-axis direction, z represents the z-axis direction, and u y represents the center of the Gaussian function in the y direction, u z represents the center of the Gaussian function in the z direction, σ y represents the variance of the distribution in the y direction, σ z represents the variance of the distribution in the z direction;

[0049] Characteristic variables, including time t, spatial position x, and particle size d, are extracted from the fitting results to form a characteristic matrix.

[0050] It is worth pointing out that the X-axis refers to the front of the human body, the Y-axis is the direction perpendicular to the X-axis in the plane, and the Z-axis refers to the direction perpendicular to the plane formed by the X-axis and the Y-axis in space. Since coughing is sprayed along a certain direction (in front of the human body) and has a Gaussian distribution law, a model of cough propagation in space and time is constructed by fitting at different times. Therefore, in this embodiment, fitting a two-dimensional Gaussian function with the concentration distribution in the space formed by the X-direction grid and the Y-axis and Z-axis can be closer to the actual situation and more in line with actual laws.

[0051] Specifically, in this example, since a fine grid unit consisting of 20×80×160 grids is constructed in the three-dimensional calculation domain of 4×4×3m, the N droplet distribution databases specifically cover 156,600 spatiotemporal distribution Gaussian function models to describe 11 different sizes of droplet particles in the indoor environment. These models span nine spatial positions along the x-axis, with an interval of 0.2m and a time range of 0.5s to 15s (time interval of 0.01s, a total of 1450 time steps). Specifically, between 5-60μm, and within the range of 5μm intervals, there are actually 11 particles that are easy to spread and easy to cause infection, so 11 different sizes of droplet particles are described in this embodiment. In this embodiment, the X-axis direction is divided into 20 grids. In the actual propagation process, affected by the distance, the particles from the first grid are likely to be transmitted to the 2nd-10th grids, so this embodiment selects nine spatial positions along the x-axis. This is only an example and not a limitation. In other feasible implementations, the data on the type and spatial position of the particles can be adjusted within a certain range according to actual conditions. However, no matter how it is changed, it is within the scope of protection of this application.

[0052] Optionally, the S4 includes:

[0053] The feature matrix is ​​used as the input of the machine learning model, and the spatial position (x, y, z) of the particle at time t is used as the output to train the model, and the N droplet distribution databases are integrated into a unified target Gaussian model. The input of the target Gaussian model is time t, spatial position x, and particle size d, and the output is the corresponding a, y, z, uy 、u z , σ y and σ z .

[0054] In this optional embodiment, the spatial position (x, y, z) of the particle at a specific time t is considered based on the Adaptive polynomial random forest algorithm (APRF). The characteristic variables t, x and d (size) and the output variables of the Gaussian function parameters are obtained through the Gaussian model. These variables are processed in parallel to construct a comprehensive feature matrix as input and output data of the machine learning model.

[0055] When solving the target Gaussian model, a hyperparameter optimization strategy combining Bayesian grid search and cross-validation accurately determined the optimal combination of the maximum polynomial degree, the number of decision trees (n_estimators), and the maximum depth of decision trees (max_depth), minimizing the root mean square error (RMSE) of the model on the validation set.

[0056] Specifically, the dataset was divided into training and test sets in a ratio of 7:3, with a random seed of 42 to ensure repeatability. Evaluation on an independent test set showed that for the existing dataset of 10μm to 60μm (interval 5μm), the maximum MAE was 0.0015 and the maximum MSE was 0.0036. Expanding the range to include position data from 10μm to 60μm, the maximum MAE and MSE increased to 0.0035 and 0.0078, respectively. Further expanding the range to 5μm and 65μm outside the data range, the maximum MAE and MSE reached 0.1608 and 0.03922, respectively, verifying the high accuracy of the model in predicting the spatiotemporal distribution of particles of different sizes using MSE and MAE.

[0057] Below, an experiment is used to describe the above droplet distribution prediction method based on adaptive polynomial random forest as follows:

[0058] Take cough droplets with a particle size of 18um as an example for prediction calculation, in a room of 4×4×3m, as follows Figure 2 The indoor distribution concentration of 18um cough droplets in the present invention is a calculation example, including an indoor model, a cough droplet inlet (diameter 0.02m), and a cough height of 2m.

[0059] The database working condition results of the temporal and spatial distribution of cough droplets of different particle sizes of 10-60um (5um interval) in 0-15s are simulated by numerical calculation, such as Figure 3A fine grid space consisting of 20×80×160 grids is constructed in the computational domain of 4×4×3m to independently calculate and analyze the concentration of particles.

[0060] Taking the Gaussian fitting of the spatial distribution concentration of 10μm, 30um, and 50um droplet cough particles in the range of 0.4-0.6m in the X direction in 6s as an example, the data samples corresponding to the working conditions are shown in Table 1. The two-dimensional Gaussian function fitting distribution corresponding to the 10um droplet is as follows Figure 4-5 shown.

[0061] Table 1 Sample numerical calculation conditions

[0062]

[0063] Based on the APRF algorithm, the spatial position (x, y, z) of the particle at a specific time t is used. The characteristic variables t, x and d, as well as the output variables of the Gaussian function parameters are obtained through the Gaussian model. The algorithm is as follows Figure 6 shown.

[0064] The present invention selects a particle size of 18 μm as a representative for prediction, revealing the concentration distribution of cough droplets with a cough time of 6 seconds in the x-axis direction of 0.8-1.0 m, such as Figure 7 shown.

[0065] The results show that the peak values ​​at different locations are consistent with the predicted concentrations from numerical simulations, with an error within ±5%, verifying the accuracy of the model in the core area, e.g. Figure 8 shown. Figure 8 The predicted concentration and original concentration obtained by cutting along the y and z directions are shown. The center position is close to the actual value, accurately reflecting the particle distribution dynamics. Although there are differences in the data intervals at different times, the actual data are very close to the predicted line, especially in the peak area, showing excellent prediction performance.

[0066] In summary, through the rapid prediction method of indoor cough droplet distribution based on the APRF algorithm proposed by the present invention, the diffusion of 50 kinds of droplets of different particle sizes in different time and space can be quickly predicted through the calculation conditions of 11 kinds of cough droplets of different particle sizes in 0-15s. For this embodiment, it takes about 789 minutes to numerically calculate the 61 total conditions of different particle sizes in different time and space, while the rapid prediction method of indoor cough droplet distribution based on the APRF algorithm proposed by the present invention only takes about 185 minutes to calculate the droplet two-dimensional Gaussian function model based on the APRF algorithm. Compared with the traditional numerical calculation method, the proposed rapid prediction method of indoor cough droplet spatiotemporal distribution based on the APRF algorithm can save about 77% of the calculation time, which can significantly improve the calculation efficiency and response time of droplet diffusion distribution, greatly reduce the calculation cost, and provide risk response in time.

[0067] The embodiment of the present application also provides a droplet distribution prediction system based on an adaptive polynomial random forest, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program. The droplet distribution prediction system based on an adaptive polynomial random forest can implement various embodiments of the above method and achieve the same beneficial effects, which will not be described in detail here.

[0068] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A droplet distribution prediction method based on adaptive polynomial random forest, characterized in that: include: S1: setting droplet simulation conditions, and simulating droplet distribution based on the simulation conditions; S2: constructing grid cells based on the simulation conditions, and independently calculating and analyzing the number and concentration of particles in each grid cell; S3: performing two-dimensional Gaussian function fitting based on the number and concentration of the particles, and constructing N droplet distribution databases based on the fitting results, each of the droplet distribution databases including Gaussian fitting parameters, particle size, and time and space; S4: Obtaining a target Gaussian model based on the N droplet distribution databases; S5: Solving and verifying the target Gaussian model to obtain a final droplet distribution prediction model, and predicting droplet distribution according to the real-time cough time, direction and size based on the droplet distribution prediction model.

2. The droplet distribution prediction method based on adaptive polynomial random forest according to claim 1 is characterized in that: The S1 includes: In the fluid mechanics simulation software, set the preset spatial range; The particle size range of the droplets simulated based on the Euler-Lagrange model within the preset spatial range is N1μm to N2μm, with an interval of nμm, and the distribution of cough droplets within a set time period under a set ambient temperature and a set relative humidity.

3. The droplet distribution prediction method based on adaptive polynomial random forest according to claim 1 is characterized in that: The S2 includes: Determine a computational domain based on the simulation conditions, and divide the computational domain into a plurality of grid units according to a preset division condition; Count and analyze the number and concentration of particles in each grid cell; The preset division conditions are as follows: The numerical stability in each grid cell is higher than a set stability threshold, and the grid density is higher than a set density threshold, wherein the density threshold is determined based on whether subtle changes in droplets can be captured.

4. The droplet distribution prediction method based on adaptive polynomial random forest according to claim 3 is characterized in that: The method of calculating and analyzing the number and concentration of particles in each grid cell includes: In each time step, the number and concentration of cough droplets in each grid cell are recorded, and the concentration data at all times are summarized to form a time series data set of concentration distribution, and the concentration distribution of cough droplets in the X, Y, and Z directions is recorded.

5. The droplet distribution prediction method based on adaptive polynomial random forest according to claim 1 is characterized in that: The S3 includes: A two-dimensional Gaussian function is fitted to the concentration distribution of the space composed of an X-direction grid and the Y-axis and Z-axis with 1s as a moment, and N droplet distribution databases are obtained, which satisfy the following relationship: In the formula, a represents the amplitude, y represents the y-axis direction, z represents the z-axis direction, and u y represents the center of the Gaussian function in the y direction, u z represents the center of the Gaussian function in the z direction, σ y represents the variance of the distribution in the y direction, σ z represents the variance of the distribution in the z direction; Characteristic variables, including time t, spatial position x, and particle size d, are extracted from the fitting results to form a characteristic matrix.

6. The droplet distribution prediction method based on adaptive polynomial random forest according to claim 5 is characterized in that: The S4 includes: The feature matrix is ​​used as the input of the machine learning model, and the spatial position (x, y, z) of the particle at time t is used as the output to train the model, and the N droplet distribution databases are integrated into a unified target Gaussian model.

7. A droplet distribution prediction system based on adaptive polynomial random forest, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method described in any one of claims 1 to 6 are implemented.

Images