Intelligent prediction and sterilization linkage system for workshop environment microbial sedimentation risk

Abstract

The invention relates to the technical field of workshop environment, in particular to an intelligent prediction and sterilization linkage system for workshop environment microbial settlement risk, which comprises a risk prediction module used for performing fusion calculation on multi-dimensional environment data and dynamic behavior data acquired in a workshop to obtain a microbial settlement risk map, and performing sterilization and sterilization on the microbial settlement risk map; identifying a high-risk area in the microbial sedimentation risk map, and generating a killing instruction according to the high-risk area; the risk disinfection and killing module is used for executing directional disinfection and killing operation on the high-risk area according to the received disinfection and killing instruction, and updating the microbial sedimentation risk map based on feedback data after the disinfection and killing operation is completed, so that the problem that microbial risk prediction and disinfection and killing control of an existing workshop environment are separated from each other, and the risk of the workshop environment is influenced is solved. And intelligent closed-loop linkage based on real-time risk perception cannot be formed.

Description

Technical Field

[0001] This invention relates to the field of workshop environment technology, specifically to an intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in workshop environments. Background Technology

[0002] In industrial workshops with strict requirements for cleanliness, such as those in the food, pharmaceutical, and precision electronics industries, microbial sedimentation is a key risk source affecting product quality and safety. Therefore, effective prediction and timely disinfection of the risk of microbial sedimentation in the workshop environment can realize the transformation from passive response to proactive prevention and control, which is of great significance for preventing pollution, reducing product scrap rate, and ensuring production safety.

[0003] Currently, common methods for controlling microbial activity in workshop environments typically rely on regular manual sampling and testing combined with sensors at fixed locations for environmental monitoring. They also involve starting and stopping broad-spectrum disinfection equipment such as UV lamps and sprayers based on preset thresholds or fixed schedules, forming an independent monitoring and execution system.

[0004] However, as mentioned above, the two core links of risk prediction and disinfection execution in the workshop environment are disconnected from each other, failing to form an intelligent closed loop of "perception-decision-execution-verification". This results in blind and inefficient disinfection operations, making it impossible to achieve precise risk control and optimal resource allocation. Summary of the Invention

[0005] To address the technical problem that existing workshop environments suffer from fragmented microbial risk prediction and disinfection control, which prevent the formation of an intelligent closed-loop linkage based on real-time risk perception, this application provides an intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in workshop environments.

[0006] The intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in a workshop environment provided in this application adopts the following technical solution: A smart prediction and disinfection linkage system for the risk of microbial sedimentation in a workshop environment includes: The risk prediction module is used to fuse and calculate the multi-dimensional environmental data and dynamic behavior data collected in the workshop to obtain a microbial sedimentation risk map, identify high-risk areas in the microbial sedimentation risk map, and generate disinfection instructions based on the high-risk areas. The risk disinfection module is used to perform targeted disinfection operations on high-risk areas according to the received disinfection instructions, and to update the microbial sedimentation risk map based on the feedback data after the targeted disinfection operation is completed.

[0007] Furthermore, the steps for fusing and calculating the multidimensional environmental data and dynamic behavioral data collected in the workshop to obtain the microbial sedimentation risk map include: The spatiotemporal environmental features related to microbial activity and diffusion are extracted from multidimensional environmental data, and dynamic behavioral data are identified. The intensity and scope of the impact of dynamic behavioral data on microbial disturbance are quantitatively evaluated to obtain disturbance quantification data. Based on the timestamps and spatial coordinates corresponding to the dynamic behavior data and the perturbation quantization data, the dynamic behavior data and the perturbation quantization data are mapped to a preset spatiotemporal grid coordinate system to obtain spatiotemporally aligned dynamic behavior grid data and perturbation quantization grid data. Then, the weight values ​​on each spatiotemporal grid in the preset spatiotemporal grid coordinate system are calculated to form a dynamic behavior data weight field and a perturbation quantization data weight field. Based on the weight field of dynamic behavior data and the weight field of perturbation quantization data, the dynamic behavior grid data and the perturbation quantization grid data are weighted and summed to obtain fused data; Risk projection is performed based on the fused data to generate a map representing the risk of microbial sedimentation in future time periods.

[0008] Furthermore, the steps for extracting spatiotemporal environmental features related to microbial activity and diffusion from multidimensional environmental data include: Spatiotemporal alignment and normalization are performed on the multidimensional environmental data to obtain a multidimensional environmental data sequence; Physical field calculations were performed based on multidimensional environmental data sequences to obtain the microbial growth potential field calculated from temperature and humidity, the particle concentration field calculated from particulate matter concentration values, and the airflow field calculated from wind speed and direction values. Spatiotemporal correlation analysis and feature fusion were performed on the microbial growth potential field, particle concentration field and airflow field to extract a high-dimensional fusion feature tensor; Based on the high-dimensional fusion feature tensor, the weights of different spatiotemporal locations and different physical field features on microbial activity images are calculated. Then, the high-dimensional fusion feature tensor is weighted, aggregated, and dimensionality reduced based on the weights to generate spatiotemporal environmental features.

[0009] Furthermore, the steps to identify dynamic behavioral data, quantify and assess the intensity and scope of impact of dynamic behavioral data on microbial disturbance, and obtain quantitative disturbance data include: Based on the behavior recognition mapping table, behavior category events are identified from dynamic behavior data, and event feature vectors are extracted from each behavior category event; The perturbation intensity is calculated based on the event feature vector to obtain the quantified value of the perturbation intensity for each behavior category event; Based on the quantified value of the disturbance intensity and the spatial location corresponding to each behavioral event, the propagation and attenuation of the air disturbance generated by each behavioral event in the workshop space are simulated to obtain the disturbance influence field of each behavioral event. The disturbance influence field is spatiotemporally superimposed and fused to obtain disturbance quantification data.

[0010] Furthermore, the steps for generating a microbial sedimentation risk map representing future time periods based on the fused data include: The fused data is fed into the initial risk model for processing to obtain the initial risk field characterizing the spatial distribution of microbial sedimentation risk at the current moment; By extrapolating the evolution rules of the initial risk field over future periods based on the preset diffusion-convection attenuation mechanism, a future risk evolution sequence is obtained; Based on the preset risk level mapping table, the future risk intensity at each moment in the future risk evolution sequence is mapped to the risk level, and rendered according to the spatiotemporal coordinates to generate a microbial sedimentation risk map.

[0011] Furthermore, the steps for identifying high-risk areas on the microbial sedimentation risk map and generating disinfection instructions based on these high-risk areas include: In the microbial sedimentation risk map, risk areas that exceed the preset risk intensity and meet the preset threat conditions are marked as high-risk areas, and the spatial range characteristics and risk intensity characteristics of high-risk areas are extracted. Based on the spatial characteristics of each high-risk area and the status of pre-installed disinfection equipment, corresponding disinfection equipment is allocated to each high-risk area, and disinfection intensity is allocated to each high-risk area based on the risk intensity characteristics of each high-risk area, and then disinfection instructions are generated.

[0012] Further steps for extracting the spatial extent and risk intensity characteristics of high-risk areas include: Geometric and topological analysis is performed on the boundary and location information of high-risk areas to extract geometric and topological features. Intensity field analysis is also performed on the risk intensity of high-risk areas to extract intensity statistical and intensity distribution features. By integrating geometric and topological features, spatial extent features are obtained; and by integrating intensity statistical features and intensity distribution features, risk intensity features are obtained.

[0013] Furthermore, the steps for carrying out targeted disinfection operations in high-risk areas based on received disinfection instructions include: The received disinfection instructions are parsed to obtain the device identifier of the target disinfection equipment, the coordinates of the target disinfection area, the target disinfection dosage, and the action parameters; Based on the equipment identification and the coordinates of the target disinfection area, path and attitude planning is performed. Combined with the workshop environment map and the status of the disinfection equipment, the motion path and actuator attitude of the target disinfection equipment to reach and cover the high-risk area are generated. Based on the motion path, actuator posture, target disinfection dosage, and action parameters, the target disinfection equipment is driven to perform targeted disinfection operations on high-risk areas.

[0014] Beneficial effects achieved: This application provides an intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in a workshop environment, comprising: a risk prediction module, used to fuse and calculate multidimensional environmental data and dynamic behavioral data collected in the workshop to obtain a microbial sedimentation risk map, identify high-risk areas in the microbial sedimentation risk map, and generate disinfection instructions based on the high-risk areas; and a risk disinfection module, used to perform targeted disinfection operations on the high-risk areas according to the received disinfection instructions, and update the microbial sedimentation risk map based on the feedback data after the targeted disinfection operations are completed.

[0015] In this application, the risk prediction module integrates multi-dimensional environmental data and dynamic behavioral data collected from the workshop to generate a microbial sedimentation risk map reflecting the spatial distribution of risks. Based on this map, high-risk areas are directly identified, and specific disinfection instructions are generated. This logically couples risk perception and decision-making for the first time, solving the problem of the separation between prediction and control at the instruction level. Next, the risk disinfection module receives and executes the disinfection instructions generated by the risk prediction module, realizing targeted disinfection operations on high-risk areas. The disinfection actions are directly derived from the risk perception results, achieving precise connection from risk information to execution actions and solving the problem of blind execution. Finally, the risk disinfection module updates the microbial sedimentation risk map based on the feedback data after the targeted disinfection operation is completed. This allows the disinfection effect of the previous round to affect the risk prediction of the next round in real time, enabling self-adjustment and optimization based on the execution results. Thus, a dynamically adjusted and continuously iterative intelligent closed-loop linkage is formed in the entire chain of "perception-decision-execution-verification". This fundamentally integrates the traditionally independent prediction and disinfection units into an adaptive and optimizable intelligent overall system. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in the workshop environment of this application.

[0017] Explanation of icon numbers: 10. Risk prediction module; 20. Risk elimination module. Detailed Implementation

[0018] The following combination Figure 1 This application will be described in further detail.

[0019] This application discloses an intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in workshop environments.

[0020] Please refer to Figure 1 The intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in the workshop environment proposed in this embodiment includes: The risk prediction module 10 is used to fuse and calculate the multi-dimensional environmental data and dynamic behavior data collected in the workshop to obtain a microbial sedimentation risk map, identify high-risk areas in the microbial sedimentation risk map, and generate disinfection instructions based on the high-risk areas; the risk disinfection module 20 is used to perform targeted disinfection operations on the high-risk areas according to the received disinfection instructions, and update the microbial sedimentation risk map based on the feedback data after the targeted disinfection operation is completed.

[0021] By constructing a closed-loop system tightly integrated with the risk prediction module 10 and the risk disinfection module 20, automation and intelligence are achieved from risk perception to precise intervention. This integrates the previously isolated risk prediction and disinfection processes into a coherent data-driven workflow. The risk prediction module 10 is responsible for fusing multi-dimensional environmental data and dynamic behavioral data to generate a visualized microbial sedimentation risk map and, based on this, generate disinfection instructions, thus providing clear objectives and justifications for disinfection actions. The risk disinfection module 20, on the other hand, is responsible for translating the disinfection instructions generated by the risk prediction module 10 into specific targeted physical interventions, i.e., executing targeted disinfection operations.

[0022] After the targeted disinfection operation is completed, feedback data from multiple sources will be collected. This mainly includes changes in monitoring values ​​of environmental sensors (such as particulate counters and specific gas sensors) deployed in high-risk areas and their surrounding environment before and after disinfection; confirmation of dosage delivery and status logs transmitted back by the target disinfection equipment itself; and subsequent periodic microbial sedimentation sampling or online rapid detection results. After spatiotemporal alignment, this feedback data will be used to synchronously collect time-series data from sensors deployed in high-risk areas and the environment before and after the targeted disinfection operation, obtaining the baseline risk intensity value (e.g., colony forming units CFU / m³) for each environmental sensor location before disinfection. 3The risk reduction rate (i.e., the ratio of the previous value minus the subsequent value to the previous value) is calculated for each point based on the concentration of particulate matter and the risk intensity value during the stabilization period after disinfection. Simultaneously, disinfection dose data for each target disinfection device is extracted from its logs, showing the doses applied to high-risk areas during operation. This dose data is then mapped to the same spatial coordinate system as the environmental sensor grid through the movement path and actuator posture of the target disinfection device, forming a dose spatial distribution map. Next, using the dose value of each environmental sensor point as the independent variable and the risk reduction rate of that point as the dependent variable, a spatial weight matrix is ​​calculated for each environmental sensor point using a Gaussian kernel function or a double-squared kernel function, based on the spatial coordinates (such as latitude and longitude or two-dimensional coordinates within the workshop). This spatial weight matrix defines the influence weights of all other environmental sensor points on the currently calculated point (hereinafter referred to as the target point). The value decays with increasing spatial distance to ensure that closer points contribute more during local fitting. For each target point, the system uses the aforementioned weight matrix to weight the observed datasets of dose values ​​and risk reduction magnitude, centered on that target point, to construct a local linear regression model. The local regression coefficients (including intercept and slope) of the target point are solved using the weighted least squares method. Specifically, the system minimizes the weighted sum of squared residuals, i.e., it solves for the parameter estimate that minimizes the weighting error, thereby obtaining the quantitative relationship function between the dose value and the risk reduction magnitude at the target point. Then, the local fitting process is repeated for all target points to obtain the set of local regression coefficients for each spatial location. The set of local regression coefficients is then interpolated and smoothed in space to generate a continuous dose-effect response surface. The value of each point on the dose-effect response surface characterizes the strength of the influence of the dose at the location of the environmental sensor on risk reduction.

[0023] Using the dose-response surface derived above, interpolation calculations are performed on the entire workshop environment covered by the microbial sedimentation risk map. The dose values ​​from the dosage logs stored on the target disinfection equipment are then input into the dose input response function reflected by the dose-response surface. Where D is the dose value, These are points obtained from the dose-response surface. Position (horizontal axis is) The vertical axis is The effect coefficient of the function, the output of the function. This represents the risk reduction amount at the corresponding point, ultimately generating a risk reduction effect field corresponding to the microbial sedimentation risk map. Each pixel value in the risk reduction effect field represents the degree of risk intensity reduction caused by the targeted disinfection operation at that location.

[0024] By performing pixel-by-pixel subtraction between the two risk mitigation effect fields with identical spatial resolution and the microbial sedimentation risk map, and subtracting the value of the risk mitigation effect field from the value of the microbial sedimentation risk map to obtain the residual field, optimization algorithms such as gradient descent are used to automatically adjust the key function parameters used to calculate the effective sedimentation probability in the initial risk model, such as parameters describing the natural decay rate of microorganisms, the contact inactivation efficiency of pesticides, or the surface adhesion coefficient, with the goal of minimizing the sum of squared residuals over the entire region. This ensures that the predicted mitigation field output by the adjusted initial risk model approximates the risk mitigation effect field as closely as possible, thereby making the initial risk model more accurate in predicting risk mitigation under similar future conditions.

[0025] After fine-tuning the parameters of the initial risk model, the updated initial risk model, combined with the latest multidimensional environmental data and dynamic behavior data, is used to re-execute the calculation process from feature fusion to risk inference, generating a revised microbial sedimentation risk map. This transforms the traditional passive, periodic, or fixed-threshold-based targeted disinfection operations into a proactive, real-time targeted disinfection mode guided by the microbial sedimentation risk map. The system can not only detect where the risk is, but also drive the corresponding equipment to accurately eliminate it, verify the elimination effect, and learn to optimize subsequent judgments. This significantly improves the timeliness, accuracy, and resource utilization efficiency of workshop microbial risk control, achieving adaptive, closed-loop intelligent management of microbial sedimentation risk.

[0026] In one feasible implementation, the specific execution steps of the risk prediction module include steps S11 to S16: Step S11: Extract spatiotemporal environmental features related to microbial activity and diffusion from multidimensional environmental data, identify dynamic behavioral data, quantify and evaluate the intensity and scope of influence of dynamic behavioral data on microbial disturbance, and obtain disturbance quantification data.

[0027] This step enables the construction of two types of dynamic behavioral data and perturbation quantification data that are both independent and intrinsically related for subsequent risk fusion calculations. This overcomes the shortcomings of existing methods that rely solely on a single data source or simply mix and process multiple data sources, resulting in one-sided risk causal analysis and ambiguous physical meaning.

[0028] By employing two parallel processing paths, the system explores the causes of risks from two dimensions: the workshop environment and human activities. On the one hand, it extracts spatiotemporal environmental features from multidimensional environmental data of the workshop environment, aiming to transform raw sensor data such as temperature, humidity, particulate matter, and airflow into physical field features that can directly characterize the growth potential and spatial diffusion ability of microorganisms, thereby revealing the inherent risk tendencies of the workshop environment itself. On the other hand, it identifies dynamic behavioral data and obtains disturbance quantification data, aiming to parse activities such as personnel movement and equipment operation from video or sensor signals into event-based quantitative descriptions of the intensity and spatial impact range of microbial disturbances. This captures the dynamic driving factors that trigger instantaneous changes in risk, providing the system with feature inputs that are both physically interpretable and computationally processable. This allows the system to not only understand how workshop environmental conditions are conducive to the survival and spread of microorganisms, but also to clearly identify "which specific behaviors, where, and at what intensity disturbed the environment and may have caused risks," thus laying a solid data foundation for achieving more accurate and granular risk prediction and tracing.

[0029] It should be noted that in the workshop environment, acquiring multidimensional environmental data is achieved by deploying an integrated sensor network. This sensor network consists of various types of sensor nodes arranged in a spatial topology along key areas and flow paths. These include temperature and humidity sensors for collecting temperature and humidity data, laser particle counters for monitoring the concentration of suspended particulate matter of different sizes, multi-directional anemometers for measuring airflow velocity and direction, and gas sensors for detecting the concentration of volatile organic compounds or specific microbial metabolic markers. These sensor nodes are connected to a data aggregation gateway via wired or wireless industrial IoT protocols to achieve high-frequency synchronous acquisition of data at the second or sub-second level. The data is then uniformly timestamped and labeled with device location coordinates, forming multidimensional environmental data with spatiotemporal consistency. Simultaneously, to acquire macroscopic environmental parameters and calibration data, auxiliary sensors capable of measuring atmospheric pressure, illuminance, and noise are typically deployed at key locations in the workshop. All sensors are connected to a central data platform for preprocessing, such as time synchronization, outlier filtering, and missing value interpolation, thereby constructing multidimensional environmental data with high spatiotemporal resolution for characterizing microbial activity and diffusion potential.

[0030] Furthermore, step S11 may also include steps S111 to S118: Step S111: Perform spatiotemporal alignment and normalization on the multidimensional environmental data to obtain a multidimensional environmental data sequence.

[0031] First, spatiotemporal alignment is based on the spatial coordinates of each sensor node and the timestamp of its data acquisition. Environmental data from different locations, which may have different sampling times and frequencies, are uniformly mapped to a preset three-dimensional spatial grid coordinate system covering the entire workshop and a unified high-frequency time series through spatial interpolation and temporal resampling techniques. This ensures that every moment and every spatial grid point has multi-dimensional environmental parameter values ​​that strictly correspond to their spatial location.

[0032] Specifically, the process involves: pre-setting a three-dimensional spatial grid coordinate system covering the entire workshop environment and a unified high-frequency reference time series. Spatially, for each moment, the environmental parameter observations reported by all sensor nodes and their known three-dimensional spatial coordinates are first acquired. Then, based on these discrete spatial point data, a semi-variogram model is constructed according to the spatial positional relationships between all known sampling points to quantify the autocorrelation and range of environmental parameters within the workshop environment. Next, the optimal weights for each interpolated spatial grid point in the three-dimensional spatial grid coordinate system are calculated using this semi-variogram model. Finally, by weighted summing of the observations from all known sampling points, the estimated environmental parameter values ​​at the corresponding spatial grid points are calculated. By traversing and calculating each spatial grid point in the three-dimensional spatial grid coordinate system, a continuous and smooth environmental parameter field covering the entire workshop environment is generated, thus transforming discrete point data into continuous spatial data usable for subsequent full-field analysis.

[0033] In the time dimension, for each spatial grid point in the three-dimensional spatial network coordinate system, its time series is initially composed of the results of spatial interpolation at each time step. However, due to the different original sampling times of different sensors, the time series is discontinuous at the initial time step. Therefore, time resampling is required. That is, based on a unified high-frequency reference time series, the existing observations of each spatial grid point at the original sampling time step are used to calculate the parameter values ​​of the spatial grid point at each unified reference time point using methods such as linear interpolation or spline interpolation. This generates a continuous and equally spaced environmental parameter time series at a unified timestamp for each spatial grid point.

[0034] After the above spatial interpolation and temporal resampling processes, all spatial grid points and environmental parameters of all dimensions are regularized in spatiotemporal order to obtain a multidimensional environmental data sequence that is spatially continuous and temporally synchronized, providing a strictly consistent spatiotemporal data foundation for subsequent feature extraction and fusion analysis.

[0035] Next, normalization is performed on each dimension of environmental data that has been spatiotemporally aligned. This is done by Z-score standardization based on the statistical characteristics (such as historical mean and historical standard deviation) of the historical data of the corresponding dimension, or by minimum-maximum scaling based on its physical range. This transforms the values ​​of all dimensions into a unified dimensionless standard value range. For example, for each dimension of environmental data that needs to be processed, the historical mean and historical standard deviation of the historical data in the corresponding dimension are calculated. For each environmental parameter at a spatial grid point, the historical mean of the corresponding dimension is subtracted and then divided by the historical standard deviation of the corresponding dimension to obtain the standardized value of each spatial grid point in the corresponding dimension. This transforms the original values ​​of all environmental parameters into a set of unified dimensionless standard values ​​with a mean of 0 and a standard deviation of 1, thereby mapping the data of all dimensions to the same scale range.

[0036] After the above two steps, the resulting multidimensional environmental data sequence is a structured numerical data that is strictly aligned in the spatiotemporal dimension and standardized in the numerical scale. It can effectively eliminate the data heterogeneity problem caused by differences in sensor deployment locations, asynchronous sampling, and different dimensions and numerical ranges of various physical quantities. It provides clean, consistent, and directly mathematically operable input data for all subsequent feature extraction and risk prediction based on data fusion and deep learning.

[0037] The semi-variogram model mentioned above is a mathematical model used to quantify the spatial autocorrelation of regionalized variables. In the context of workshop environmental parameter Kriging interpolation, it specifically refers to a function used to characterize the variation of environmental parameters such as temperature and humidity within the three-dimensional space of the workshop. This semi-variogram model fits a mathematical function by calculating the relationship between the square of the difference in environmental parameter values ​​between all known sensor sampling point pairs and the spatial distance between these point pairs. Its typical form is a spherical model or an exponential model. The graph of this function clearly shows the spatial structure characteristics of the parameters: when two points are very close, their environmental parameter values ​​are usually similar, and the semi-variogram value is small; as the distance increases, the similarity weakens, and the function value increases until it reaches a relatively stable plateau (called the sill value); the spatial distance corresponding to the function value growing from the initial value to the sill value is called the range, which directly defines the effective range of influence of the environmental parameter with significant autocorrelation within the workshop space.

[0038] Step S112: Perform physical field calculations based on the multidimensional environmental data sequence to obtain the microbial growth potential field calculated from temperature and humidity, the particle concentration field calculated from particulate matter concentration values, and the airflow field calculated from wind speed and direction values.

[0039] ① For the microbial growth potential field, the standardized values ​​of temperature and humidity at each spatial grid point are read and input into a preset microbial growth response function. ,in, The actual observed value representing the ambient temperature; The actual observed value representing the relative humidity of the environment; and These represent the minimum and maximum temperatures at which the target microorganism can grow, respectively, defining its growth temperature range; It is a curve shape parameter that controls the temperature as it approaches the maximum temperature. The steepness of the decline in growth rate; This represents the minimum relative humidity threshold required for microbial growth; below this value, growth is considered inhibited. This is a humidity-affecting factor used to determine the sensitivity of humidity to transitioning from an inhibitory state to a growth-promoting state, i.e., when... equal hour, The value is exactly 0.5; This represents a growth rate constant; when temperature or humidity conditions do not meet the above range, the microbial growth potential field... The value is 0. This microbial growth response function quantifies the potential rate of microbial reproduction under given temperature and humidity conditions, thereby outputting a microbial growth potential field that covers the entire workshop environment and can characterize growth potential.

[0040] ② For particle concentration fields, the standardized particulate matter concentration values ​​are typically categorized into several classes (e.g., PM2.5, PM5, PM10) based on aerodynamic particle size. Each particle size class is assigned a weighting coefficient reflecting its relative ability to carry and transport microorganisms; generally, larger particle sizes have higher weighting coefficients. Then, the standardized concentration values ​​for each particle size class at each spatial grid point are multiplied by their corresponding weighting coefficients and summed to obtain a weighted particulate matter concentration index. Finally, this weighted particulate matter concentration index is multiplied by a preset linear transformation coefficient. ,in The weighted particulate matter concentration index representing the i-th historical sample. This represents the microbial sedimentation concentration measured under the same spatiotemporal location and conditions, or the reference concentration of carrier particles as determined by authoritative methods. and Representing all historical samples respectively and The mean, calculated The value is the linear transformation coefficient, which is directly mapped to a scalar value representing the equivalent concentration of carrier particles that can carry microorganisms. By performing the same calculation on all spatial grid points, a continuously distributed carrier particle concentration field covering the entire workshop environment is generated.

[0041] ③ For the airflow field, read the wind speed (i.e., speed magnitude) and wind direction angle (usually defined clockwise with true north as the 0-degree reference) at each spatial grid point. First, record the wind speed value at each spatial grid point. and wind direction angle values Convert to vector components in a Cartesian coordinate system, for example, using the formula and Calculate the east and west directions respectively. (East is positive) and north-south direction ( The wind speed component (north is positive) is used to express the wind speed value and wind direction angle value of each spatial grid point as a two-dimensional horizontal vector. , Next, the two-dimensional horizontal vectors on all spatial grid points are organized and stored in space, forming a discrete vector data set covering the entire workshop environment. To obtain a continuous spatial representation, the cell in the two-dimensional horizontal grid where any location is situated is typically determined, i.e., the coordinates of its four nearest known spatial grid points and their corresponding two-dimensional horizontal vectors are found. Then, for... The components are first calculated along the x-direction (or longitude direction) for the two left and two right grid points. The values ​​are linearly interpolated separately to obtain two intermediate values. Then, linear interpolation is performed on these two intermediate values ​​in the y-direction (or latitude direction) to obtain the estimated value of the U component at that location. The same interpolation process is repeated for the V component to obtain the estimated value of the V component at that location. After obtaining the U and V components at any location, the wind speed at that point is calculated by combining them. Angle of wind direction The angle returned by the atan2 function is usually expressed in mathematical standard, i.e., starting from due east as 0 degrees and increasing counterclockwise. Therefore, if you need to convert it to a wind direction angle defined as due north as 0 degrees and increasing clockwise, you need to perform the corresponding coordinate transformation, for example... This process synthesizes the wind speed and wind direction at that point. Through this series of processes, a continuously distributed airflow field that can describe the direction and speed of airflow at any location in the workshop is directly constructed.

[0042] In this step, environmental monitoring data is transformed into several core dynamic fields that directly drive the diffusion, transport, growth, and sedimentation of microorganisms through physical mechanisms, laying a physical foundation for the subsequent fusion of these fields and the deduction of the dynamic propagation of microorganisms.

[0043] Step S113 involves performing spatiotemporal correlation analysis and feature fusion on the microbial growth potential field, particle concentration field, and airflow field to extract a high-dimensional fusion feature tensor.

[0044] The microbial growth potential field, particle concentration field, and airflow field are aligned in the same spatiotemporal grid coordinate system and used as different input channels to form a multi-channel spatiotemporal data cube. This spatiotemporal data cube is then input into a three-dimensional convolutional neural network (3D convolutional neural network). This 3D convolutional neural network automatically learns and extracts the interaction features between the microbial growth potential field, particle concentration field, and airflow field by sliding calculations within a local spatiotemporal neighborhood (e.g., considering a spatial region and several time steps before and after it simultaneously) through its multi-layer 3D convolutional kernels. For example, one convolutional kernel of the 3D convolutional neural network may learn to identify "high growth potential, high particle concentration, and airflow field". The high-risk combination pattern of "upstream airflow field" can be captured by another convolutional kernel, which may learn to capture the diffusion pattern of "rapid propagation of particle concentration along a specific airflow direction." These local correlation features extracted by the primary convolutional layer are further combined and abstracted through deeper convolutions and nonlinear activation functions in the three-dimensional convolutional neural network. Finally, they are integrated and reshaped into a high-dimensional fusion feature tensor at the output layer of the three-dimensional convolutional neural network, such as a four-dimensional tensor with the shape (96, 50, 50, 64). The first dimension 96 represents the time series length of 4 consecutive days with one time step per hour, and the second and third dimensions 50 and 50 respectively represent the discretization of the workshop environment plane into 50 rows and 50 columns. The spatial grid points, with the fourth dimension 64 representing 64 different fusion feature channels learned and extracted through a three-dimensional convolutional neural network, have a specific value in this four-dimensional tensor. For example, the value at location (t=48, x1=25, y1=30, c=15) is 0.87. Its physical meaning can be interpreted as follows: at the midpoint of the monitoring time series (the 48th hour, i.e., 12 noon on the 3rd day), the spatial location at the workshop environmental plane coordinates (25, 30) has an activation intensity of 0.87 for a specific risk pattern characterized by the 15th feature channel of the three-dimensional convolutional neural network. This pattern is "under moderate humidity conditions, accompanied by a specific southeast wind direction, fine..." The quantitative expression of the spatiotemporal correlation feature of "the product of particulate matter concentration and microbial growth potential reaching a local peak" encapsulates information on multiple complex risk patterns at all spatiotemporal locations in the tensor. This tensor can be directly used as input to downstream deep learning models (such as risk prediction classifiers). In this way, the microbial growth potential field, particle concentration field, and airflow field, which respectively represent the "breeding conditions," "propagation carriers," and "diffusion dynamics" of microorganisms, are deeply integrated into a comprehensive risk feature expression that can fully represent the complete propagation chain of "microbial growth - carrier carrying - airflow diffusion" through a data-driven approach. This provides a high-quality input containing complex spatiotemporal correlation information for subsequent accurate risk prediction.

[0045] Step S114: Based on the high-dimensional fusion feature tensor, calculate the weights of different spatiotemporal locations and different physical field features on the microbial activity image, and perform weighted aggregation and dimensionality reduction on the high-dimensional fusion feature tensor based on the weights to generate spatiotemporal environmental features.

[0046] The high-dimensional fusion feature tensor is input into a spatial attention submodule and a channel attention submodule. The spatial attention submodule performs global average pooling on the input high-dimensional fusion feature tensor along the channel dimension, averaging all feature channel values ​​at each spatiotemporal location. This compresses the channel dimension, resulting in a feature map that retains only spatiotemporal information. Its dimension is the same as the time step and planar grid dimension of the input high-dimensional fusion feature tensor. This feature map is then input into a lightweight network consisting of two convolutional layers. The first convolutional layer uses a small kernel for feature transformation to capture local spatiotemporal dependencies. The second convolutional layer outputs a single-channel feature map, and the output value at each location is mapped to between 0 and 1 using a sigmoid activation function, serving as the attention weight for that spatiotemporal location. This process is repeated for each spatiotemporal location, resulting in an attention weight map with the same spatial dimension (i.e., time and planar grid dimension) as the input high-dimensional fusion feature tensor. This attention weight map quantifies the differences in the importance of different spatiotemporal locations on microbial activity.

[0047] The channel attention submodule performs global average pooling on each feature channel (i.e., the feature channels derived from the microbial growth potential field, particle concentration field, and airflow field) of the input high-dimensional fusion feature tensor. This involves calculating the average value of the feature value of the channel at all spatiotemporal grid points, thereby compressing the global spatiotemporal information of each feature channel into a scalar to obtain an initial channel description vector. The channel description vector is then input into an activation network consisting of two fully connected layers. The first fully connected layer reduces the channel dimension to capture the nonlinear dependencies between feature channels, followed by a nonlinear transformation introduced by a ReLU activation function. The second fully connected layer restores the channel dimension to its original dimension. Finally, a Sigmoid activation function normalizes each element of the output vector to between 0 and 1, forming a channel attention weight vector. This attention weight vector quantifies the differences in the contribution of different feature channels to the influence of microbial activity.

[0048] The attention weight map output by the spatial attention submodule is expanded to a four-dimensional shape identical to the high-dimensional fusion feature tensor by adding a dimension or through a broadcast operation, ensuring alignment of its temporal, height, and width dimensions and a channel dimension of 1. Simultaneously, the attention weight vector output by the channel attention submodule is also expanded to the same four-dimensional shape through reshaping and broadcast operations, ensuring alignment of its channel dimensions and a spatiotemporal dimension of 1. Then, these two expanded attention weights are directly multiplied or sequentially weighted element-wise with the values ​​at each spatiotemporal location and feature channel in the high-dimensional fusion feature tensor. The weight values ​​range from 0 to 1; when a weight value is close to 1, the corresponding value is preserved or enhanced; when a weight value is close to 0, the corresponding value is suppressed or weakened. This achieves synergistic enhancement of key spatiotemporal locations and important feature channels while automatically suppressing non-critical information.

[0049] The weighted high-dimensional fusion feature tensor is input into a fully connected layer or a 1x1 convolutional layer for linear transformation and channel dimension compression. If a fully connected layer is used, all spatiotemporal dimensions of the weighted high-dimensional fusion feature tensor are flattened into a one-dimensional vector, and then a linear transformation is performed through a weight matrix. The number of output neurons in this weight matrix is ​​set to be less than the dimension of the weighted high-dimensional fusion feature tensor, thus directly achieving channel dimension compression and global feature fusion. If a 1x1 convolutional layer is used, there is no need to flatten the weighted high-dimensional fusion feature tensor; a 1x1 convolutional kernel is applied independently at each spatiotemporal location. The number of channels in this convolutional kernel is the same as the number of channels in the weighted high-dimensional fusion feature tensor, but the number of output channels is set to be fewer. Convolution operations are performed on each spatiotemporal location... The feature channels of a location are linearly combined to significantly reduce the number of channels while maintaining the spatiotemporal dimension, thereby compressing local features and integrating cross-channel information. Regardless of the method used, the output after linear transformation will introduce nonlinearity through an activation function (such as ReLU), resulting in a spatiotemporal environmental feature with lower dimensionality but more condensed and focused information. This highlights the regions that have a significant impact on microbial activity in time and space, as well as the key feature combinations extracted from the microbial growth potential field, particle concentration field, and airflow field. At the same time, irrelevant or redundant information is downplayed, so that the generated spatiotemporal environmental feature can more accurately and efficiently characterize the core environmental state most relevant to microbial risk, thereby significantly improving the accuracy and interpretability of subsequent risk prediction models.

[0050] Step S115: Based on the behavior recognition mapping table, identify behavior category events from the dynamic behavior data, and extract event feature vectors from each behavior category event.

[0051] It should be noted that the behavior recognition mapping table is a predefined rule base or a trained lightweight classifier that corresponds to dynamic behavior data and behavior categories. It contains behavior thresholds or pattern templates for distinguishing different high-risk behaviors, such as "people gathering", "material unpacking", "equipment cleaning" and "rapid crossing".

[0052] Dynamic behavior data is segmented into time segments, and key features (such as changes in the number of people in the area, movement trajectories, specific action postures, and object interaction states) are extracted from each time segment. These key features are then matched with rules in a behavior recognition mapping table or input into a classifier to determine which behavior category the behavior occurring within that time segment belongs to. A behavior category label, start time, end time, and spatial coordinates are then assigned to each behavior category event. Next, event feature vectors are extracted from each behavior category event. For each behavior category event, the corresponding values ​​(but not limited to) the duration of the behavior category event, the average number of people or objects involved, the area of ​​the core region, the average speed of moving parts, the intensity of the action (such as changes in acceleration), and the spatial coordinates of the event's center point are arranged in a fixed order and normalized to form a numerical vector representing the dimensional characteristics of that behavior category event—the event feature vector.

[0053] Step S116: Calculate the perturbation intensity based on the event feature vector to obtain the quantized value of the perturbation intensity for each behavior category event.

[0054] The event feature vector is input into a pre-trained disturbance intensity assessment model, which is usually a lightweight fully connected neural network or a support vector regression machine. Its internal parameters have been trained with a large amount of historical data and have learned the nonlinear mapping relationship between the characteristics of different behavioral categories of events, such as the type of behavioral category, duration, number of people involved, and movement speed, and the disturbance intensity of air and surface microorganisms.

[0055] The perturbation intensity assessment model feeds the input event feature vector into the first fully connected layer. This first fully connected layer has a weight matrix and a bias vector. The input event feature vector is linearly combined and transformed through weight matrix multiplication and bias vector addition to generate the first-layer hidden feature vector. Then, the first-layer hidden feature vector is processed by a non-linear activation function (such as ReLU) to introduce the non-linear expressive power required by the perturbation intensity assessment model. Then, the processed first-layer hidden feature vector is fed into one or more subsequent fully connected layers. Each layer repeats a similar process of linear transformation, bias addition, and non-linear activation, thereby abstracting and integrating the complex patterns in the input features layer by layer. Finally, the features processed by all hidden layers are passed to the output layer, which is a single-neuron fully connected layer without an activation function or using a linear activation function. It directly maps the final high-dimensional feature representation to a single continuous value through a final linear weighted sum. This value is the perturbation intensity quantification value of the event of that behavior category.

[0056] Step S117: Based on the quantified value of the disturbance intensity and the spatial location corresponding to each behavioral category event, simulate the propagation and attenuation of the air disturbance generated by each behavioral category event in the workshop space to obtain the disturbance influence field of each behavioral category event.

[0057] The quantified disturbance intensity value and spatial location are input into a parameterized air disturbance propagation model. This model is typically based on a simplified atmospheric diffusion equation or a Gaussian plume / puff model. Its core is to define a function of spatial location. This function calculates the spatial diffusion distribution and exponential decay of the quantified disturbance intensity value from that behavioral event category in space, based on the distance and direction from the event's spatial location to the target grid point (possibly modified by incorporating the background airflow field), and a preset turbulent diffusion coefficient. By calculating for each spatial grid point in the three-dimensional spatial grid of the workshop environment, a scalar field covering the entire workshop environment, with the highest intensity near the spatial location, and gradually decaying to zero with increasing distance, can be obtained. This scalar field represents the disturbance influence field of that single behavioral event category. This transforms each discrete point-like behavioral event category into a continuous spatial influence distribution based on physical laws, thereby quantitatively characterizing the potential range and intensity gradient of the impact of that behavioral event category on the risk of microbial sedimentation at different locations within the workshop environment.

[0058] Step S118: Spatiotemporal superposition and fusion of the disturbance influence field to obtain disturbance quantification data.

[0059] Define a unified time reference point (usually the current moment), and apply a preset time decay function (such as an exponential decay function) to weight the quantized value of the disturbance intensity of the corresponding disturbance field based on the time difference between the occurrence time of the behavioral category event corresponding to each disturbance influence field and the event reference point, so as to obtain the influence intensity value, in order to simulate the process of the disturbance effect gradually weakening over time.

[0060] All time-decay-weighted disturbance influence fields are aligned point-by-point according to their spatial grid coordinates, and each disturbance influence field is mapped onto the 3D spatial grid of the workshop environment. This ensures that each disturbance influence field has a corresponding disturbance intensity quantization value at each spatial grid point. Then, for the multiple behavioral category events to be fused, the influence intensity value corresponding to each behavioral category event, after being weighted by the time-decay function, is read sequentially. For each spatial grid point, the disturbance influence fields of all behavioral category events are traversed, and the influence intensity value of each disturbance influence field at the corresponding spatial grid point is extracted. Influence intensity values ​​belonging to the same spatial location but from different behavioral category events are added together; the sum is the quantized value of the corresponding spatial grid point after fusion. The total disturbance intensity is calculated by repeatedly performing this traversal and summation operation on all spatial grid points, ultimately generating a three-dimensional data field that can cover the entire workshop environment. The value of each spatial grid point in this three-dimensional data field represents the cumulative disturbance intensity from all behavioral events at the corresponding spatial location, which is the behavioral disturbance quantification data. It can characterize the real-time cumulative spatial disturbance intensity distribution of all behavioral events in the workshop environment on the microbial environment at the current moment after considering the time decay effect. This integrates the impact of behavioral events from multiple discrete times and different spatial locations into a continuous, unified, and timely accurate disturbance quantification data, thereby providing dynamic disturbance input for subsequent fusion with environmental feature data and risk prediction.

[0061] Step S12: Based on the timestamps and spatial coordinate information corresponding to the dynamic behavior data and the perturbation quantization data respectively, the dynamic behavior data and the perturbation quantization data are mapped to a preset spatiotemporal grid coordinate system to obtain spatiotemporally aligned dynamic behavior grid data and perturbation quantization grid data. Then, the weight value on each spatiotemporal grid in the preset spatiotemporal grid coordinate system is calculated to form a dynamic behavior data weight field and a perturbation quantization data weight field.

[0062] Based on the timestamps and spatial coordinate information carried by dynamic behavioral data and perturbation quantization data, these two types of data are mapped to a preset spatiotemporal grid coordinate system and a high-frequency time series.

[0063] Specifically, for dynamic behavioral data, each behavioral category event is assigned to a corresponding time frame based on its timestamp. A preset spatiotemporal grid coordinate system covering the entire workshop's three-dimensional space and continuous time axis is established, where space is divided into uniform cubic grids and time is divided into continuous equal-length segments (time frames). When a behavioral category event occurs, its timestamp is read, it is assigned to the corresponding specific time frame, and then its spatial coordinates are processed. These spatial coordinates are typically three-dimensional point coordinates (such as the position of the center of a person's torso). If the nearest neighbor interpolation method is used, the Euclidean distance from the event coordinate point to the center of all grid points in the preset spatial grid is calculated, and the nearest spatial grid point is found. Then, the entire event feature vector of the event is directly assigned to this nearest spatial grid point, while other spatial grid points do not have data for this event within that time frame. If the inverse distance weighted interpolation method is used, a spatial influence radius is set with the event coordinate point as the center, all spatial grid points within this radius are found, the distance from the event point to the center of each grid point is calculated, and the data is processed according to the distance... The reciprocal (or other decay function) assigns a weight to each grid point (the closer the distance, the greater the weight). Then, the value of each feature dimension of the event feature vector is multiplied by the weight and accumulated to the corresponding feature dimension of the corresponding grid point. This achieves the diffusion and distribution of the features of a behavior category event to multiple grids around it. Regardless of the method used, this process is repeated for all behavior category events within their respective time frames. In the end, each spatiotemporal grid (i.e., a specific spatial grid under a specific time frame) may contain some or all feature vector information from one or more behavior category events. After these information are aggregated (such as summation or averaging), dynamic behavior grid data that is strictly aligned with the preset spatiotemporal grid coordinate system in both time and space dimensions is formed.

[0064] For perturbation-quantized data that is already in the form of a spatial field, a unified time series with a fixed time interval is preset. The perturbation-quantized data itself is a spatiotemporal data field that has been discretized into a grid in space, but may have irregular timestamps or different sampling rates in time. During alignment, each target time frame (e.g., the Tth second) in the unified time series is traversed, and a search and match is performed on the time axis of the perturbation-quantized data. If the timestamp of the perturbation-quantized data at a certain spatiotemporal grid point is completely consistent with the target time frame, the perturbation-quantized data at that timestamp is directly assigned to the grid point of the target frame. If the timestamps are not completely matched, a time-dimensional resampling interpolation algorithm is used, such as nearest neighbor interpolation. That is, for each target time frame, the data frame with the closest timestamp in the perturbation-quantized time series is found, and the grid values ​​of its entire spatial field are directly copied to the target frame. The above time interpolation calculation is performed independently on each spatiotemporal grid point in the spatiotemporal data field, thereby generating a complete perturbation intensity field with a consistent spatial grid structure for each unified target time frame. Through this process, the perturbation quantization data, which may have been non-uniform in time, is regularized to a unified time axis shared with the dynamic behavior grid data, and finally forms perturbation quantization grid data that is strictly aligned with the preset spatiotemporal grid coordinate system in both time and space dimensions.

[0065] Next, the weight values ​​on each spatiotemporal grid are calculated to form two weight fields. This step analyzes the instantaneous quality and reliability indicators of the two types of data at each spatiotemporal grid point in real time (e.g., the local sampling density and confidence of dynamic behavior grid data, and the spatial smoothness and gradient stability of perturbation quantization grid data). Through a lightweight neural network or a set of heuristic rules, the weight ratio of dynamic behavior data and perturbation quantization data at each grid point in the subsequent fusion is dynamically calculated, and the weight fields of dynamic behavior data and perturbation quantization data are output respectively, ensuring that their contribution is enhanced in areas with high data quality and their influence is suppressed in areas with sparse data or high noise.

[0066] Step S13: Based on the weight field of dynamic behavior data and the weight field of perturbation quantization data, perform weighted summation on the dynamic behavior grid data and the perturbation quantization grid data to obtain fused data.

[0067] After reading four values ​​from each spatiotemporal grid—the dynamic behavior grid data value, the perturbation quantization grid data value, the dynamic behavior data weight value of the dynamic behavior data weight field, and the perturbation quantization data weight value of the perturbation quantization data weight field—the dynamic behavior grid data value is multiplied by the dynamic behavior data weight value to obtain the weighted dynamic behavior data value. Simultaneously, the perturbation quantization grid data value is multiplied by the perturbation quantization data weight value to obtain the weighted perturbation quantization value. Then, the weighted dynamic behavior data value and the weighted perturbation quantization value are added together, and the sum is the fused data on the corresponding spatiotemporal grid. By traversing each spatiotemporal grid in the spatiotemporal grid coordinate system and repeating the above weighted multiplication and addition operation, the fused data on each spatiotemporal grid is obtained.

[0068] Step S14: Based on the fused data, risk projection is performed to generate a map representing the risk of microbial sedimentation in future time periods.

[0069] By transforming the fused data corresponding to each spatiotemporal grid into decision-making information, the limitations of traditional monitoring systems, which can only reflect the current or past state and cannot predict the development trend of risks, are overcome. By performing forward spatiotemporal simulation calculations on the workshop environment situation represented by the current fused data, the evolution trend and spatial distribution of microbial sedimentation risk in the future can be deduced. The output microbial sedimentation risk map not only reveals the existing high-risk areas, but also predicts in advance when, where, and with what intensity the risk will generate or spread, thus shifting the decision point for risk prevention and control from post-event response to pre-event warning.

[0070] Furthermore, step S14 may also include steps S141 to S143: Step S141: Input the fused data into the initial risk model for processing to obtain the initial risk field characterizing the spatial distribution of microbial sedimentation risk at the current moment.

[0071] It should be noted that the initial risk model is a computational function or algorithm module that encapsulates the basic physical mapping relationship between microbial sedimentation and the current workshop environmental state, receiving fused data as its sole input. This initial risk model processes all fused data at the current moment as a whole. Its internal logic parses the multidimensional environmental data contained in the fused data at each spatiotemporal grid point, and based on built-in transformation rules, such as those based on sedimentation probability, the initial risk model first substitutes the air disturbance intensity represented by the input fused data at the corresponding spatiotemporal grid into a Sigmoid function: Where C represents the fused data value of the current grid point, C0 is an empirical threshold parameter (representing the critical disturbance intensity that begins to significantly affect sedimentation), and k is a sensitivity coefficient controlling the slope of the curve. When the local disturbance is very weak (C is much smaller than C0), the airflow is stable, microbial particles are easily suspended, the sedimentation probability P is low, and the output approaches 0. As the air disturbance intensity increases and crosses the critical point C0, the probability of microbial particles colliding with the surface increases, and the sedimentation probability rises rapidly. When the air disturbance intensity is extremely strong, the sedimentation probability approaches the upper limit of 1, indicating that sedimentation almost inevitably occurs. The parameters C0 and k are jointly calibrated by combining computational fluid dynamics simulation data and experimental observation data of microbial sedimentation in real environments to ensure that the mapping relationship conforms to physical reality.

[0072] Through the above mapping process, the air disturbance intensity on each spatial grid is assigned a settlement probability with a clear physical interpretation. At the same time, a basic risk coefficient related to spatial location is invoked. This basic risk coefficient is determined by the inherent properties of the corresponding spatiotemporal grid (such as whether it is close to critical equipment or whether it is a resident area). The effective settlement probability coefficient is multiplied by the basic risk coefficient to obtain a risk value that represents the absolute risk. In order to facilitate subsequent comparison and deduction, methods such as range normalization are used to scale the corresponding risk values ​​on all spatiotemporal grids to a dimensionless exponential interval, thereby obtaining the initial risk field.

[0073] Step S142: deduce the evolution rules of the initial risk field in the future period based on the preset diffusion convection attenuation mechanism to obtain the future risk evolution sequence.

[0074] It should be noted that the pre-defined diffusion-convection attenuation mechanism describes three physical processes of microbial sedimentation risk in the workshop environment: the diffusion mechanism simulates the spontaneous diffusion of microbial sedimentation risk from high-concentration areas to low-concentration areas due to air turbulence; the convection mechanism simulates the process by which microbial sedimentation risk substances are carried and transported by macroscopic airflow fields (such as ventilation airflow); and the attenuation mechanism simulates the effect of natural reduction of microbial sedimentation risk due to particle sedimentation, surface adsorption, and disinfection. The corresponding equation is: in, Risk field representing the risk value of microbial sedimentation at spatial location x and time t; It is the rate of change of the risk field over time; the three terms on the right side of the equation correspond to the three preset physical mechanisms, namely the first term. It is a convection mechanism that describes the risk field. How to be affected by velocity field (Transportation from real-time monitoring or simulated airflow fields); Item 2 It is a diffusion mechanism that describes the diffusion of risk due to turbulent mixing, and its intensity is determined by the turbulent diffusion coefficient. Control; Third item It is a decay mechanism that describes the exponential decay of risk due to the combined effects of sedimentation, adsorption, and deactivation, and its rate is determined by the decay rate constant. Decide.

[0075] In the simulation, the initial risk field is used as the initial condition for the calculation. Real-time air velocity and wind direction data acquired from the environmental monitoring system are used as the velocity field input for the convection term. The turbulent diffusion coefficient and comprehensive attenuation rate, pre-calibrated according to the site characteristics, are used as fixed parameters. Numerical methods (such as the finite volume method) are employed to iteratively solve the above equations on a discrete spatiotemporal grid with short time steps, thereby determining the initial risk field at the current moment. The simulation is designed to predict a series of future moments. risk field This process constitutes a sequence of future risk evolutions, generating a video documenting how future risks move, spread, and attenuate with airflow. This provides a basis for decision-making in predicting the movement path, spread range, and intensity changes of high-risk areas in advance, achieving a leap from passive monitoring to proactive prediction.

[0076] Step S143: Based on the preset risk level mapping table, the future risk intensity at each moment in the future risk evolution sequence is mapped to the risk level, and rendered according to the spatiotemporal coordinates to generate a microbial sedimentation risk map.

[0077] It should be noted that the preset risk level mapping table is a reference standard table that divides continuous risk values ​​into several discrete risk levels (such as low, medium, and high). For example, risk values ​​in the range [0, 0, 3) can be set as low risk, [0.3, 0.7) as medium risk, and [0.7, 1.0] as high risk.

[0078] During processing, the system reads the future risk evolution sequence, which contains the risk values ​​of each spatiotemporal grid in the entire spatial region at multiple consecutive future moments. By traversing each spatial grid in the risk field at each moment, the risk value within that spatiotemporal grid is extracted. Then, according to a preset risk level mapping table, the interval in which the risk value falls is queried, thus converting it into a corresponding risk level label. After completing the risk level mapping for all spatiotemporal grids at all moments, the system renders the data according to spatiotemporal coordinates, generating a risk spatial distribution map for each future moment. Each spatial grid in the map is filled with a corresponding preset color according to its assigned risk level, such as green for low risk, yellow for medium risk, and red for high risk. The series of images arranged in chronological order are then combined to form a microbial sedimentation risk map. This microbial sedimentation risk map can be a dynamic visualization video or a set of static snapshots arranged in a timeline, allowing managers or automated systems to clearly identify when, where, and what level of risk will occur in the future, providing the most direct and clear graphical decision-making basis for accurate early warning and intervention.

[0079] Step S15: In the microbial sedimentation risk map, risk areas that exceed the preset risk intensity and meet the preset threat conditions are marked as high-risk areas, and the spatial range features and risk intensity features of the high-risk areas are extracted.

[0080] From the microbial sedimentation risk map representing the overall risk situation, the most urgent and intervention-required core targets, i.e., high-risk areas, are selected. This transforms macroscopic risk visualization information into specific and actionable disinfection targets, solving the problem of vague disinfection targets and scattered resources caused by the inability to automatically identify high-risk areas in traditional methods. By setting a dual judgment logic of preset risk intensity and preset threat conditions, the microbial sedimentation risk map is intelligently analyzed. Specifically: It should be noted that the preset risk intensity is a pre-set risk value threshold, such as initially screening areas with a risk value greater than 0.7 as potential high-risk areas; while the preset threat conditions are used to further identify high-risk areas that truly pose a threat from the potential high-risk areas, including but not limited to the length of time that the risk intensity of the potential high-risk area continues to exceed the preset risk intensity, the area expansion rate of the potential high-risk area, or whether its spatial location covers predefined key sensitive points such as operating tables or food processing lines.

[0081] During runtime, the microbial sedimentation risk map is scanned frame by frame to identify all sets of risk values ​​that exceed the preset risk intensity. Then, preset threat conditions are applied to these sets for verification. Only areas that simultaneously meet the preset risk intensity and preset threat conditions will be judged and marked as high-risk areas.

[0082] Furthermore, step S15 may also include steps S151 to S152: Step S151: Perform geometric and topological analysis on the boundary and location information of high-risk areas to extract geometric and topological features, and perform intensity field analysis on the risk intensity of high-risk areas to extract intensity statistical features and intensity distribution features.

[0083] Regarding the extraction of geometric features, a starting pixel is selected within the high-risk region. Then, according to the rules of the Moore's Neighborhood Tracking algorithm, starting from this starting pixel, the search continues in a clockwise or counterclockwise direction within its eight neighborhoods to find the next pixel belonging to the region boundary. This region boundary is usually defined as the position where the pixel inside the high-risk region is adjacent to the external background pixel. Each time a boundary pixel is found, its coordinates are recorded, and the search continues with this boundary pixel as the new center to find the next neighboring boundary pixel. This process is iterated until the tracking path returns to the starting pixel, thus forming a closed sequence of boundary pixel coordinates. The total number of pixels within the region corresponding to the boundary pixel coordinate sequence is counted to obtain the pixel count area. The Euclidean distances between adjacent boundary pixels in the boundary pixel coordinate sequence are summed to obtain the perimeter. The perimeter is divided by the pixel count area to obtain the perimeter-area ratio, which quantifies the shape complexity of the region corresponding to the boundary pixel coordinate sequence. A higher ratio usually indicates a more tortuous boundary. The obtained geometric features include pixel count area, perimeter, and perimeter-area ratio.

[0084] Regarding the extraction of topological features, the connectivity between multiple high-risk areas is analyzed to determine whether they form a single threat or multiple isolated points. The minimum Euclidean distance between the boundaries of each pair of high-risk areas is calculated, or the DBSCAN algorithm is used for processing. For example, with the DBSCAN algorithm, the geometric center of each high-risk area is used as a data point. A neighborhood distance threshold (such as a pixel distance of 5) and a minimum number of points (such as 2) are set. The core points of areas that are spatially close and densely connected are grouped into the same cluster. Thus, all high-risk areas belonging to the same cluster are determined to be spatially connected and merged into a single threat, while isolated areas that cannot be grouped into any cluster are determined to be isolated points. Then, for each high-risk area, the coordinates of the geometric center (or centroid) are determined by calculating the arithmetic mean of the coordinates of all pixels within it. Then, the Euclidean distance (using the distance formula between two points) and the relative azimuth angle between the center point and each preset key location (such as an air vent or workbench) are calculated. The relative azimuth angle is usually calculated with true north as the 0-degree reference. The angle between the direction of the line connecting the center point and the key location and true north is calculated using the arctangent function. The resulting topological features include the connectivity between high-risk areas, the Euclidean distance to the preset key locations, and the relative azimuth angle.

[0085] Secondly, within the same spatial range, intensity field analysis is performed on the risk values ​​corresponding to high-risk areas to extract intensity statistical features. This includes accessing the risk intensity stored in each spatial pixel within the high-risk area to form an intensity data set of risk intensity values; summing the risk intensities of all pixels in this set and then dividing by the total number of pixels in the area to obtain the average intensity; traversing the intensity data set and finding the maximum value through comparison to obtain the peak intensity; calculating the intensity standard deviation, which is used to calculate the intensity difference between the risk intensity value of each pixel and the average intensity, squaring each intensity difference and summing them, then dividing by the total number of pixels to obtain the variance, and finally taking the square root of the variance to obtain the intensity standard deviation. The obtained intensity statistical features include: average intensity, peak intensity, intensity difference, and intensity standard deviation.

[0086] Furthermore, intensity distribution features are extracted, and gradient calculations are performed on the risk intensity field within the high-risk area. Typically, gradient operators such as Sobel or Prewitt from image processing are used for convolution operations to obtain the intensity change rate of each pixel in the east-west (X) and north-south (Y) directions, thus constructing the intensity gradient vector for that pixel. The dominant intensity gradient direction of the entire high-risk area is determined by principal component analysis or by directly calculating the average direction of the gradient vectors of all pixels within the area (by calculating the direction angle of the sum of all vectors). This direction (e.g., with due east as 0 degrees) indicates the main spatial direction in which the risk spreads or attenuates outward from the high-intensity core. The spatial heterogeneity index of intensity is quantified by calculating a spatial autocorrelation statistic such as the global Moran's index. The calculation process is as follows: First, a spatial weight matrix (usually based on pixel adjacency or the inverse of distance) is defined among pixels within a high-risk region. Then, based on the intensity risk intensity of all pixels, the deviation from the global mean is calculated. The Moran's index value is then derived using a formula (involving the sum of the product of the spatial weight matrix and the deviation value). This index value ranges from -1 to 1. A positive value indicates a positive spatial correlation in risk intensity (i.e., clustering of high or low values, exhibiting patchy clustering), a negative value indicates a negative correlation (interleaving of high and low values), and a value close to 0 indicates a random spatial distribution (i.e., uniform distribution). The resulting intensity distribution features include the intensity gradient vector, the intensity gradient direction, and the intensity spatial heterogeneity index.

[0087] Step S152: Combine geometric features and topological features to obtain spatial range features, and combine intensity statistical features and intensity distribution features to obtain risk intensity features.

[0088] Geometric and topological features are respectively regularized into one-dimensional numerical arrays. Then, these two arrays are sequentially concatenated along the feature dimension to form a spatial extent feature. This spatial extent feature encodes both the inherent morphology of the high-risk area and its relative positional relationship with the environment.

[0089] Simultaneously, the extracted intensity statistical features and intensity distribution features are concatenated using the same method to form a risk intensity feature. This risk intensity feature describes the overall risk level, internal volatility, and spatial variation pattern of the high-risk area. This integrates multiple scattered feature indicators that originally belonged to different analytical dimensions into two structured feature representations, providing a refined and comprehensive input for the next step of disinfection strategy planning. This enables precise disinfection decisions to be made based on a feature profile that simultaneously includes the target's "location, size, shape, and relationship with key points" as well as "average danger, location of the most dangerous point, risk variation, and clustering." For example, for a high-risk area that is large, complex in shape, close to key workstations, and has highly concentrated internal risks with a clear direction of diffusion, a customized encirclement disinfection and targeted suppression strategy that perfectly matches these complex features can be generated.

[0090] Step S16: Based on the spatial range characteristics of each high-risk area and the status of the pre-set disinfection equipment, allocate corresponding disinfection equipment to each high-risk area, and based on the risk intensity characteristics of each high-risk area, allocate corresponding disinfection intensity to each high-risk area, and then generate a disinfection instruction.

[0091] The spatial range characteristics of each high-risk area are matched and calculated with the status of the pre-set disinfection equipment. Appropriate disinfection equipment is assigned to each area. For example, mobile disinfection robots with path planning capabilities are given priority to high-risk areas with large areas and irregular shapes. The same wide-area spraying equipment with adjustable angle is assigned to multiple adjacent and connected small high-risk areas. The nearest mobile equipment is assigned to high-risk areas that are far from the fixed equipment or located in blind spots. This achieves the optimal matching of equipment and target high-risk areas in terms of spatial coverage.

[0092] Simultaneously, based on the risk intensity characteristics of each high-risk area, corresponding disinfection intensities are assigned to each high-risk area. Specifically, a preset intensity-parameter mapping table is used to convert the numerical values ​​corresponding to the risk intensity characteristics into specific disinfection operation parameters. For example, the average intensity and peak intensity are mapped to the unit area spray dosage and action time of the disinfectant, and the intensity gradient direction is mapped to the key spray direction of the equipment end effector, thereby generating a customized disinfection intensity strategy for each high-risk area. Finally, the assigned equipment identifier, the spatial coordinates of the target high-risk area, and the calculated target disinfection dosage are integrated and encapsulated into a disinfection instruction containing "target high-risk area - equipment identifier - target disinfection dosage". This allows the deduced risk prediction results to automatically drive the most suitable disinfection resources in the physical world to intervene in the high-risk area with the most appropriate intensity and action method, thereby realizing a closed loop of intelligent risk prediction and disinfection execution linkage, fundamentally solving the problems of disconnect between prediction and execution and extensive and inefficient disinfection operations in traditional methods.

[0093] In one feasible implementation, the specific execution steps of the risk elimination module include steps S21 to S23: Step S21: The received disinfection command is parsed to obtain the device identifier of the target disinfection device, the coordinates of the target disinfection area, the target disinfection dosage, and the action parameters.

[0094] A command parser is invoked to decode the received disinfection command. The disinfection command is essentially a structured data packet, which encapsulates data fields representing "which device", "where to", "what dosage" and "how" according to a predefined protocol.

[0095] According to the protocol, the instruction parser reads the data stream of the disinfection instruction bit by bit. By identifying specific field separators or field header identifiers, it accurately decomposes the continuous data stream and maps it to the corresponding data structure. This allows it to extract the unique device identifier (i.e., device identification) representing the target disinfection device, the coordinate set representing the spatial range of the disinfection target (i.e., the target disinfection area coordinates corresponding to the target high-risk area), the target disinfection dose representing the total amount of disinfectant to be applied (i.e., target disinfection dose), and action parameters including action time, spray pressure, and ultraviolet wavelength. This provides the operational basis for subsequent path planning and drive execution, ensuring the accuracy, reliability, and automation level of the entire predictive disinfection linkage system in the instruction transmission stage.

[0096] Step S22: Based on the equipment identifier and the coordinates of the target disinfection area, perform path and attitude planning, and combine the workshop environment map and the status of the disinfection equipment to generate the motion path and actuator attitude of the target disinfection equipment to reach and cover the high-risk area.

[0097] Based on the device identifier, the kinematic model, workspace range, and real-time status (such as current position, remaining energy, and current idle status) of the corresponding target disinfection device are retrieved from the database of preset disinfection device status. Simultaneously, based on the coordinates of the target disinfection area, the coordinates of the target disinfection area are marked as the target area to be covered in the workshop environment map containing static obstacles and passage areas. Next, path planning is performed. If the target disinfection device is a mobile robot, a search-based algorithm (such as A algorithm) or a sampling-based algorithm (such as RRT) is used to calculate a global motion path for the target disinfection device in the workshop environment map, which allows it to safely and without collisions reach the vicinity of the target area or a specified starting point from its current position. If the target disinfection device is a fixed robot with multiple degrees of freedom, its motion trajectory in the joint space or operating space is planned so that its end effector (such as a nozzle or UV lamp) moves from the standby position to the starting pose where it can begin operation.

[0098] After generating the arrival path, the coverage path and actuator attitude planning of the target disinfection equipment are further carried out. Based on the geometry of the target area (obtained from coordinate analysis) and the operating mode of the actuator (such as reciprocating scanning and spiral coverage), combined with the local details of the target area in the workshop environment map (such as whether there are temporary obstacles), a set of fine moving path point sequences is generated to ensure that the disinfectant or radiation field can completely cover every position in the target area. The corresponding actuator attitude (such as the pitch angle and yaw angle of the nozzle, and the irradiation angle of the ultraviolet lamp) is calculated for each point on the path to ensure that the direction of action is always perpendicular to the surface of the target area or optimally pointed to the core of the risk.

[0099] Step S23: Based on the motion path, actuator posture, target disinfection dosage, and action parameters, drive the target disinfection equipment to perform targeted disinfection operations on the high-risk area.

[0100] The motion path and corresponding actuator posture are converted into motor control commands (such as wheel speeds and joint angles) using specific motion control algorithms for the target disinfection equipment (e.g., PID control algorithms for the mobile chassis and inverse kinematics solvers for the robotic arm). These commands are then sent to the motion controller of the target disinfection equipment to drive its chassis and robotic arm to move along the planned motion path and adjust the end effector to the specified posture in real time. Simultaneously, based on the target disinfection dosage and action parameters, combined with the real-time movement speed and actuator posture of the target disinfection equipment, synchronous control commands are calculated and generated to adjust the output of the actuators. This ensures that the intensity, action mode, and spatial distribution of the disinfection medium (i.e., disinfectant, ultraviolet light, etc.) match the target disinfection dosage and action parameters. Finally, through timing synchronization, the motion control and dose-parameter control of the target disinfection equipment work together to complete the targeted disinfection operation in high-risk areas. This realizes the execution of disinfection commands containing multi-dimensional parameters into physical actions, ensuring that the disinfection operation is carried out on the correct spatial path, in the correct physical state, and with the correct quality of disinfection medium, ultimately achieving the goal of automated risk elimination.

[0101] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A smart prediction and disinfection linkage system for the risk of microbial sedimentation in a workshop environment, characterized in that, include: The risk prediction module is used to fuse and calculate the multi-dimensional environmental data and dynamic behavior data collected in the workshop to obtain a microbial sedimentation risk map, identify high-risk areas in the microbial sedimentation risk map, and generate disinfection instructions based on the high-risk areas. The risk disinfection module is used to perform targeted disinfection operations on the high-risk area according to the received disinfection instructions, and to update the microbial sedimentation risk map based on the feedback data after the targeted disinfection operation is completed.

2. The intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in the workshop environment according to claim 1, characterized in that, The steps for fusing and calculating the multidimensional environmental data and dynamic behavioral data collected in the workshop to obtain the microbial sedimentation risk map include: The spatiotemporal environmental features related to microbial activity and diffusion are extracted from the multidimensional environmental data, and the dynamic behavioral data is identified. The intensity and range of influence of the dynamic behavioral data on microbial disturbance are quantitatively evaluated to obtain disturbance quantification data. Based on the timestamps and spatial coordinate information corresponding to the dynamic behavior data and the perturbation quantization data respectively, the dynamic behavior data and the perturbation quantization data are mapped to a preset spatiotemporal grid coordinate system to obtain spatiotemporally aligned dynamic behavior grid data and perturbation quantization grid data. Then, the weight value on each spatiotemporal grid in the preset spatiotemporal grid coordinate system is calculated to form a dynamic behavior data weight field and a perturbation quantization data weight field. Based on the dynamic behavior data weight field and the perturbation quantization data weight field, the dynamic behavior grid data and the perturbation quantization grid data are weighted and summed to obtain fused data; Based on the fused data, risk projection is performed to generate a map representing the microbial sedimentation risk over future periods.

3. The intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in the workshop environment according to claim 2, characterized in that, The step of extracting spatiotemporal environmental features related to microbial activity and microbial diffusion from the multidimensional environmental data includes: The multidimensional environmental data is spatiotemporally aligned and normalized to obtain a multidimensional environmental data sequence. Physical field calculations are performed based on the multidimensional environmental data sequence to obtain the microbial growth potential field calculated from temperature and humidity, the particle concentration field calculated from particulate matter concentration, and the airflow field calculated from wind speed and direction. Spatiotemporal correlation analysis and feature fusion were performed on the microbial growth potential field, the particle concentration field, and the airflow field to extract a high-dimensional fusion feature tensor. Based on the high-dimensional fusion feature tensor, the weights of different spatiotemporal locations and different physical field features on the microbial activity image are calculated, and the high-dimensional fusion feature tensor is weighted, aggregated and dimensionality reduced based on the weights to generate the spatiotemporal environmental features.

4. The intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in the workshop environment according to claim 2, characterized in that, The steps of identifying the dynamic behavioral data, quantifying and evaluating the intensity and scope of the impact of the dynamic behavioral data on microbial disturbance, and obtaining quantitative disturbance data include: According to the behavior recognition mapping table, behavior category events are identified from the dynamic behavior data, and event feature vectors are extracted from each behavior category event; The perturbation intensity is calculated based on the event feature vector to obtain the quantized value of the perturbation intensity for each behavior category event; Based on the quantified value of the disturbance intensity and the spatial location corresponding to each of the aforementioned behavioral categories, the propagation and attenuation of the air disturbance generated by each of the aforementioned behavioral categories in the workshop space are simulated to obtain the disturbance influence field of each of the aforementioned behavioral categories. The disturbance influence field is spatiotemporally superimposed and fused to obtain the disturbance quantification data.

5. The intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in the workshop environment according to claim 2, characterized in that, The step of performing risk extrapolation based on the fused data to generate the microbial sedimentation risk map representing future time periods includes: The fused data is fed into the initial risk model for processing to obtain an initial risk field characterizing the spatial distribution of the microbial sedimentation risk at the current moment; By extrapolating the evolution rules of the initial risk field within the future time period based on a preset diffusion-convection attenuation mechanism, a future risk evolution sequence is obtained; Based on a preset risk level mapping table, the future risk intensity at each moment in the future risk evolution sequence is mapped to a risk level, and rendered according to spatiotemporal coordinates to generate the microbial sedimentation risk map.

6. The intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in workshop environment according to claim 1, characterized in that, The step of identifying high-risk areas in the microbial sedimentation risk map and generating disinfection instructions based on the high-risk areas includes: In the microbial sedimentation risk map, risk areas that exceed the preset risk intensity and meet the preset threat conditions are marked as high-risk areas, and the spatial range characteristics and risk intensity characteristics of the high-risk areas are extracted. Based on the spatial range characteristics of each high-risk area and the status of the pre-set disinfection equipment, corresponding disinfection equipment is assigned to each high-risk area, and based on the risk intensity characteristics of each high-risk area, corresponding disinfection intensity is assigned to each high-risk area, and then the disinfection instruction is generated.

7. The intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in the workshop environment according to claim 6, characterized in that, The steps for extracting the spatial extent and risk intensity features of the high-risk area include: Geometric and topological analysis is performed on the boundary and location information of the high-risk area to extract geometric and topological features, and intensity field analysis is performed on the risk intensity of the high-risk area to extract intensity statistical and intensity distribution features. The spatial extent feature is obtained by fusing the geometric features and the topological features, and the risk intensity feature is obtained by fusing the intensity statistical features and the intensity distribution features.

8. The intelligent prediction and disinfection linkage system for the risk of microbial sedimentation in workshop environment according to claim 1, characterized in that, The step of performing targeted disinfection operations on the high-risk area according to the received disinfection instruction includes: The received disinfection command is parsed to obtain the device identifier of the target disinfection device, the coordinates of the target disinfection area, the target disinfection dosage, and the action parameters; Based on the device identifier and the coordinates of the target disinfection area, path and posture planning is performed, and combined with the workshop environment map and the status of the disinfection equipment, the motion path and actuator posture of the target disinfection equipment to reach and cover the high-risk area are generated. Based on the motion path, the actuator posture, the target disinfection dose, and the action parameters, the target disinfection device is driven to perform the targeted disinfection operation on the high-risk area.

Images