Closed-loop intelligent thermal barrier system based on multi-mode sensing and digital twinborn verification

Abstract

The invention discloses a closed-loop intelligent thermal barrier system based on multi-modal sensing and digital twinborn verification, which belongs to the technical field of data processing and comprises a multi-modal data sensing and fusion module, an aerosol risk prediction and intelligent decision module, an intelligent thermal barrier generation and automatic deployment module and a digital twinborn verification and feedback optimization module. According to the method, the stability and robustness of state estimation can be adaptively maintained, the filtering divergence risk is reduced, the potential airflow disorder and aerosol diffusion risk is identified in advance, adaptive decision and accurate control of deployment parameters are realized, and meanwhile, quick response and reversible arrangement are realized; and the temperature field, the airflow field and the air quality in the building can be balanced and adjusted in real time. And the digital twin verification module realizes iterative optimization and parameter self-correction of a deployment scheme, so that a full-closed-loop operation link is formed. The heat insulation efficiency and the energy utilization rate are improved in the complex space environment, the pollution diffusion risk is reduced, and safe and intelligent management of the building heat environment is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a closed-loop intelligent thermal barrier system based on multi-modal perception and digital twin verification. BACKGROUND

[0002] In hospital waiting areas, fever clinics, sampling points, and ward corridors, there are often phenomena of personnel concentration, relatively closed space, and complex local heat source distribution. Research and engineering evidence shows that the thermal environment and air flow state of such spaces have an important influence on the spatial distribution and residence time of airborne pathogens (such as influenza, COVID-19, etc.) aerosols. The World Health Organization and many epidemiological studies have pointed out that respiratory infectious diseases have a significantly higher risk of aggregation in poorly ventilated and crowded places than in ordinary environments. High temperature and humidity can affect the survival cycle of pathogens, and some studies suggest that within a certain temperature and humidity range, the settling speed, floating diffusion path, and surface activity of virus aerosols will change, indirectly increasing the environmental safety pressure of medical spaces. In addition, high-temperature operating areas such as hospital logistics kitchens and disinfection supply rooms are prone to heatstroke and health risks (China CDC data in 2023: heatstroke incidence rate 12.7 / 100,000 people per year), and local high-temperature heat sources in these areas also affect local airflow structure, leading to complex phenomena such as air convection, aerosol reflux or stagnation, and increasing cross-exposure risk.

[0003] The current common environmental and prevention management methods mainly include fixed or temporary physical separation, HVAC air volume / air direction adjustment, and single measurement: for example, using tool-free assembly partitions or simple curtains for partitioning, combined with conventional body temperature screening, single-point carbon dioxide sensing, or simplified particle counting for environmental monitoring; during peak periods, manual inspection, manual air addition or subtraction, or adjustment of supply / return air opening degree are relied on to alleviate congestion and stuffiness. However, these methods have common deficiencies in engineering implementation: fixed partitions have low deployment efficiency and poor adaptability, and manual arrangement often takes more than 30 minutes per area; existing flexible thermal insulation materials have insufficient temperature resistance (mostly below 80°C), and lack the ability to deform on demand and respond in real time; single thermal imaging on the perception level is easily disturbed by environmental light or adjacent high-temperature equipment, and carbon dioxide sensors generally have a response time of more than 10s, making it difficult to reflect the instantaneous changes in air quality in a timely manner; the verification link lacks a quantitative process-result correspondence, often relying on experience for post-correction, making it difficult to form a fast and stable closed loop.

[0004] Under the premise of maintaining the order of medical treatment and preserving passage and line of sight as much as possible in medical institutions, spatial governance is shifting from a static, manual, and reactive model to a regionalized thermal-flow management model that pursues rapid, reconfigurable, and quantifiable assessment. Although the industry is showing a trend towards more intelligent thermal barrier management, practical implementation is mostly limited to single-point improvements: such as increasing sensor density in local areas, improving the trade-off between temperature resistance and light transmittance in partition materials, or introducing semi-automated movement and positioning tools at the equipment level; however, these improvements cannot provide stable and consistent control effects under conditions of strong time-varying pedestrian flow and complex airflow coupling. In this context, to build a closed-loop intelligent thermal barrier system for medical scenarios, it is necessary to introduce the joint use of multi-source heterogeneous data and on-site model verification and validation mechanisms. However, due to the strong nonlinearity and dramatic time-varying characteristics of heat-flow-crowd behavior in medical settings, the edge side is constrained by computing power and real-time requirements. If a linear approximation state estimation framework is used, it is difficult to accurately describe the actual state evolution. At the same time, process noise and observation noise will change dynamically with the environment, load and operating conditions. Fixed parameter settings are difficult to adapt in the long term, which can easily lead to abnormal amplification or rapid collapse of some principal component energy. In addition, multimodal data often have unequal sampling rates, heteroscedasticity and dimensional differences. Conventional fusion strategies are prone to causing abnormal covariance matrix spectral structure, extremely unbalanced principal component energy distribution and even loss of numerical positive definiteness, resulting in overcorrection, filter divergence or information loss during state update. Summary of the Invention

[0005] To address the aforementioned technical problems in existing technologies, embodiments of the present invention provide a closed-loop intelligent thermal barrier system based on multimodal sensing and digital twin verification. The technical solution is as follows:

[0006] The multimodal data perception and fusion module is used to collect multimodal perception data in the target space, perform spatiotemporal synchronization and coordinate alignment of the multimodal perception data, generate multimodal prediction data based on the state prediction model, and realize the fusion and recursion of multimodal perception data and multimodal prediction data through extended Kalman filtering, thereby outputting a multimodal state vector.

[0007] The aerosol risk prediction and intelligent decision-making module is used to dynamically predict the airflow and aerosol diffusion process in the target space based on multimodal state vectors and computational fluid dynamics models, construct a risk assessment model, and generate corresponding thermal barrier deployment decision instructions.

[0008] The intelligent thermal barrier generation and automatic deployment module is used to control the parameter execution and spatial arrangement of adjustable thermal barrier units with shape memory alloy skeleton and aerogel composite structure according to thermal barrier deployment decision instructions, and drive the robot cluster to automatically complete the deployment, adjustment and removal of thermal barriers to achieve spatial thermal isolation and environmental control.

[0009] The digital twin verification and feedback optimization module is used to build a digital twin scenario based on building information model and real-time multimodal perception data in a virtual simulation platform, to conduct simulation verification and effect evaluation of thermal barrier deployment decisions, and to feed back the verification results to the multimodal data perception and fusion module and the aerosol risk prediction and intelligent decision-making module to realize the automatic closed-loop optimization of the system.

[0010] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following:

[0011] 1. The closed-loop intelligent thermal barrier system based on multimodal sensing and digital twin verification provided by this invention, through the introduction of a collaborative closed-loop mechanism of multimodal data fusion, extended Kalman filter optimization, aerosol risk prediction, and digital twin verification, can achieve high-precision dynamic monitoring and intelligent control of the thermal environment of building spaces. Even under conditions of time delay, noise fluctuation, and nonlinear interference in multi-source heterogeneous sensor data, the system can adaptively maintain the stability and robustness of state estimation, reducing the risk of filter divergence caused by covariance matrix spectrum anomalies. By combining computational fluid dynamics prediction with multimodal state vectors, potential airflow turbulence and aerosol diffusion risks can be identified in advance, enabling adaptive decision-making and precise control of deployment parameters such as thermal barrier height, spacing, angle, and transmittance. Thermal barrier units based on shape memory alloy and aerogel composite structures can achieve rapid response and reversible arrangement under robot swarm drive, enabling real-time balanced regulation of the temperature field, airflow field, and air quality inside the building. Furthermore, the digital twin verification module, through dynamic simulation feedback that integrates virtual and real elements, enables iterative optimization and parameter self-calibration of the deployment scheme, thus forming a fully closed-loop operation chain of perception-prediction-decision-verification-re-optimization. In complex spatial environments, this system can improve thermal insulation efficiency and energy utilization, reduce the risk of local heat island effects and pollution spread, and achieve safe, energy-saving, and intelligent management of the building's thermal environment.

[0012] 2. This invention obtains an optimized covariance matrix through analysis, enabling dynamic identification and correction of covariance spectrum structure anomalies during multimodal sensing data fusion. This avoids numerical instability in filtering caused by excessive concentration or collapse of principal component energy, improving the accuracy and convergence of state estimation. The optimized covariance matrix obtained through analysis guarantees positive definiteness and smoothness at the numerical level, providing a stable and physically consistent mathematical foundation for subsequent calculation of the state-observation cross-covariance matrix, dynamic adjustment of the Kalman gain matrix, and recursive updates of multimodal state vectors. Therefore, the system can maintain the continuity and reliability of the filtering process even in highly nonlinear, noisy, and multi-coupled scenarios, ensuring more accurate and reliable input state data for the risk assessment model. This results in high consistency between the predicted results of thermal barrier deployment decisions and digital twin simulation verification, achieving enhanced accuracy and closed-loop steady-state operation across the entire chain from data fusion and state estimation to decision feedback.

[0013] 3. This invention obtains an optimized state-observation cross-covariance matrix through analysis, which can adaptively identify changes in the correlation between state variables and observed variables, and dynamically correct the cross-covariance relationship under abnormal disturbances or observation drift, suppressing the accumulation of bias and observation mismatch in the state estimation process. This enhances the coupling consistency between the state and observation spaces, making the subsequent Kalman gain matrix solution more robust and reliable, avoiding over-correction of filtering or response hysteresis caused by fluctuations in cross terms. Through the optimized correlation structure established at this stage, the system can achieve higher-precision recursion and fusion of multimodal state vectors, ensuring the continuous consistency of input parameters in the aerosol risk prediction model, and maintaining dynamic synchronization and parameter coordination between the virtual and real models in the digital twin simulation verification stage, ultimately achieving coordinated and stable operation of the entire closed-loop intelligent thermal barrier system at the temporal, physical, and control levels.

[0014] 4. This invention, through analysis, obtains an optimized Kalman gain matrix. This matrix dynamically adjusts the weighting of each modal information during the fusion process, enabling optimal weighting and recursive correction of the observation residuals, even under conditions of fluctuating multimodal observation data and uncertainties in state estimation. The optimized Kalman gain matrix reduces estimation oscillations caused by excessive gain or response hysteresis caused by insufficient gain, achieving a balance between accuracy and stability in the filtering update process. This results in multimodal state vectors that better reflect the actual evolution of the system, providing a highly reliable input basis for subsequent aerosol risk prediction models and making flow field simulation and risk assessment results more consistent and predictable. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a structural diagram of a closed-loop intelligent thermal barrier system based on multimodal perception and digital twin verification provided in an embodiment of the present invention.

[0017] Figure 2 This is a flowchart of the covariance matrix optimization process involved in this embodiment.

[0018] Figure 3 This is a flowchart of the optimized state-observation cross-covariance matrix analysis involved in this embodiment. Detailed Implementation

[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0020] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0021] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0022] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0023] To make the technical problem to be solved, the technical solution and advantages of the present invention clearer, a detailed description will be given below with reference to the accompanying drawings for the purpose of data analysis to predict arsenic content.

[0024] like Figure 1 The diagram shown illustrates the structure of a closed-loop intelligent thermal barrier system based on multimodal perception and digital twin verification. It includes: a multimodal data perception and fusion module, an aerosol risk prediction and intelligent decision-making module, an intelligent thermal barrier generation and automatic deployment module, and a digital twin verification and feedback optimization module.

[0025] The multimodal data perception and fusion module is used to collect multimodal perception data in the target space, perform spatiotemporal synchronization and coordinate alignment of the multimodal perception data, generate multimodal prediction data based on the state prediction model, and realize the fusion and recursion of multimodal perception data and multimodal prediction data through extended Kalman filtering, thereby outputting a multimodal state vector.

[0026] In specific embodiments, the multimodal sensing data of the target space includes, but is not limited to, real-time spatial temperature distribution data of various areas collected by devices such as infrared thermal imagers and digital temperature sensors; gas concentration sensing data collected by air quality sensors such as carbon dioxide concentration, PM2.5, PM10, and volatile organic compounds (VOCs); environmental data such as relative humidity and air pressure collected by environmental monitoring modules; airflow velocity and direction information of various points in space obtained by ultrasonic anemometers, wind vanes, or airflow field sensors based on CFD inversion; and spatial distribution, density, trajectory, and motion status data of personnel obtained by 4K visible light cameras, depth cameras, millimeter-wave radar, infrared detectors, etc.

[0027] The specific steps for spatiotemporal synchronization and coordinate alignment of multimodal sensing data are as follows:

[0028] A1 performs time alignment on multimodal sensing data from different types of sensors according to a unified system timestamp standard. It uses a master clock, NTP protocol, or hardware synchronization pulses to achieve millisecond-level time synchronization of data collected by each sensor, eliminating time errors caused by acquisition delays and clock drift in multi-source data.

[0029] A2 maps multimodal sensing data collected by various sensors to a unified coordinate system in the target space based on their physical installation location and sampling area using a coordinate transformation matrix. For images, point clouds, and vector data, methods such as extrinsic parameter calibration (e.g., calibration parameters between LiDAR and cameras) and sensor installation geometry are employed to achieve coordinate consistency of spatial data.

[0030] A3, for data with different spatial resolutions or sampling frequencies, reconstructs multimodal data to a unified spatiotemporal grid or voxel unit through spatial interpolation algorithms (such as bilinear interpolation, Kriging interpolation, etc.) and temporal resampling, which facilitates subsequent fusion processing.

[0031] A4 detects missing values, outliers, or noise in each modality of data and removes or corrects them using methods such as sliding window statistics and robust filtering. For heterogeneous modal data such as images and point clouds, radar and vision, it uses algorithms such as feature point matching and ICP (Iterative Closest Point) to achieve accurate registration of multimodal data.

[0032] A5 outputs multimodal sensing data in a unified data format after spatiotemporal synchronization and coordinate alignment are completed.

[0033] Based on the state prediction model, multimodal prediction data is generated. The specific steps are as follows:

[0034] B1. Collect historical multimodal sensing data and establish a state prediction model that can reflect the dynamic evolution law of the target space. In specific embodiments, the model can adopt physical modeling (such as heat conduction equation, CFD airflow model), statistical learning, machine learning or hybrid modeling methods.

[0035] B2 involves calibrating the parameters of the physical model or conducting supervised or unsupervised training on the data-driven model to optimize the model parameters so that they can accurately characterize the spatiotemporal evolution of each modal variable.

[0036] B3 collects and inputs multimodal sensing data in real time. Based on the pre-trained state prediction model, it recursively calculates the physical quantities of each mode in the target space at the next moment or several future moments to obtain multimodal prediction data.

[0037] Furthermore, extended Kalman filtering is used to fuse and recursively calculate multimodal sensing data and multimodal prediction data. The specific execution process is as follows:

[0038] It should be noted that the Extended Kalman Filter (EKF) is a recursive filtering algorithm widely used for state estimation of nonlinear dynamic systems. Unlike the classic linear Kalman filter, the EKF can handle cases where both the system dynamics model and the observation model are nonlinear. Its basic idea is to approximate the nonlinear system as a linear system near the current estimation point by using a first-order Taylor expansion (i.e., linearization) of the system state and observation functions at each time step. The EKF first performs prior estimation and covariance recursion of the system state using the system's nonlinear prediction model. Then, combining real-time observation data, it dynamically calculates the Kalman gain based on the linearization results of the observation model and the observation noise characteristics, and corrects and updates the state estimate and covariance. Through this prediction-correction recursive process, the EKF can effectively fuse multi-source information in complex nonlinear and noisy environments, achieving a high-precision estimate of the system's true state.

[0039] like Figure 2The diagram shows the covariance matrix optimization process in this embodiment. First, singular value decomposition is performed on the covariance matrix to obtain the eigenvalues ​​of each principal component. Then, the eigenvalues ​​of each principal component are sorted in descending order, and the cumulative energy percentage of each principal component is calculated to obtain the number of qualified principal components. Next, based on the total number of principal components and the set principal component verification percentage, the number of verified principal components is obtained, and a judgment is made: if the number of qualified principal components is greater than or equal to the number of verified principal components, the covariance matrix is ​​directly used as the optimized covariance matrix; if the number of qualified principal components is less than the number of verified principal components, spectral correction is performed, including calculating the number of qualified principal component deviations, calculating the number of additional retained principal components based on the number of deviations, and truncating the covariance matrix accordingly. Finally, the truncated covariance matrix is ​​numerically smoothed, and the numerically smoothed truncated covariance matrix is ​​used as the final optimized covariance matrix.

[0040] The covariance matrix is ​​calculated based on multimodal prediction data, and the optimized covariance matrix is ​​obtained through analysis. The specific analysis process is as follows:

[0041] In a specific embodiment, the covariance matrix is ​​calculated based on multimodal prediction data, and the specific process is as follows:

[0042] Within each prediction period, the state prediction model is used to simulate the physical quantities of each modality in the target space multiple times, and several sets of multimodal prediction data samples are obtained.

[0043] All sampled multimodal prediction data are aligned in time and space and formatted in a unified manner to ensure that the spatial points and time points corresponding to each modal physical quantity in the same set of data samples correspond one-to-one.

[0044] Each set of multimodal prediction data is concatenated into a high-dimensional feature vector in a uniform order to form a sample set (each sample is a row, and all samples form a matrix).

[0045] It should be explained that the unified order refers to the strict sorting of all multimodal prediction data according to predetermined rules for each physical quantity, spatial location, and modality category when concatenating them into a high-dimensional feature vector. This ensures that each column of all samples in matrix form has a consistent physical meaning and variable correspondence. The predetermined sorting rules are uniformly determined by the system during the design and initialization phases based on the physical quantity category, spatial layout, modality type, and sensor deployment scheme.

[0046] In specific embodiments, multimodal prediction data often involves multiple physical quantities (such as temperature, airflow velocity, humidity, CO2 concentration, etc.) and the measured values ​​of each physical quantity at multiple spatial locations. To ensure data consistency and the scientific nature of statistical analysis, the system needs to clearly define the splicing order of each physical quantity, each spatial location, and each modal category in the feature vector during the design and implementation phases, based on physical modeling, sensor layout, or business requirements. This order is usually based on the principle of prioritizing physical quantity categories, followed by increasing spatial location numbers, and can also be refined by combining device numbers, grid arrangements, or standardized field tables. For example, if the splicing order is specified as temperature at all monitoring points to airflow velocity at all monitoring points, and then CO2 concentration at all monitoring points, then each sample group should strictly arrange its components in the order of [temperature 1, temperature 2, ..., velocity 1, velocity 2, ..., CO2 concentration 1, CO2 concentration 2, ...].

[0047] The mean of the multimodal prediction data is obtained by calculating the mean of each dimension of the sample set (corresponding to the predicted value of each modal physical quantity at a certain point).

[0048] Based on the sample set, the covariance matrix among all modes and all sampling points is calculated using the covariance definition formula, specifically: Where C is the covariance matrix, X i For the i-th group of multimodal prediction samples, Let be the mean vector, i be the sample number, i = 1, 2, ..., N, and N be the number of samples.

[0049] Singular value decomposition is performed on the covariance matrix to obtain the eigenvalues ​​of each principal component. The specific operation process is as follows:

[0050] C1 takes the covariance matrix calculated from the multimodal prediction data samples as input, ensuring that it is a symmetric positive definite or semi-positive definite matrix.

[0051] It should be noted that, theoretically, the covariance matrix calculated from real sample data using the classical covariance formula is always symmetric and usually positive semi-definite (all eigenvalues ​​≥ 0). However, in practical engineering scenarios (high dimensionality, small sample size, floating-point errors, extreme distributions, etc.), it is possible for the covariance matrix to be non-strictly positive semi-definite, or even for it to contain extremely small negative eigenvalues ​​due to noise or numerical precision. Therefore, in practice, to ensure that the covariance matrix is ​​a symmetric positive definite or positive semi-definite matrix, symmetry processing and weighted regularization are used after sample calculation to ensure the symmetry of the covariance matrix and the non-negativity of all eigenvalues. If necessary, truncation or numerical smoothing is applied to the extremely small negative eigenvalues ​​in the covariance spectrum to prevent numerical ill-conditioning from affecting subsequent spectral decomposition and principal component energy analysis.

[0052] C2, Perform singular value decomposition on the covariance matrix: Where C is the covariance matrix, U is the left singular vector matrix, Σ is the singular value diagonal matrix, and V T It is the transpose of the right singular vector matrix.

[0053] It's important to explain that the left singular vector matrix is ​​an m×m orthogonal matrix, where each column vector is called a left singular vector, representing the orthogonal basis of the original data in the principal component space. The singular value diagonal matrix is ​​an m×n diagonal matrix, where the non-negative real numbers on the diagonal are called singular values, each representing variance information (energy contribution) along a principal component direction. The right singular vector matrix is ​​an n×n orthogonal matrix, where each column is a right singular vector, representing the orthogonal basis of the original data in the variable space.

[0054] It should be added that in practical engineering, almost all mainstream mathematical / engineering software (such as Python's numpy.linalg.svd, MATLAB's svd function, etc.) can obtain the transpose of the left singular vector matrix, the singular value diagonal matrix, and the right singular vector matrix in a single line of code. No manual derivation is required. Just input the covariance matrix, and the numerical library will automatically solve for all singular values ​​and their corresponding vectors.

[0055] C3 extracts all eigenvalues ​​(or singular values) from the singular value diagonal matrix, i.e., the eigenvalues ​​of each principal component. Each eigenvalue represents the energy contribution of the corresponding principal component to the overall variance.

[0056] Sort the eigenvalues ​​of each principal component from largest to smallest, and calculate the cumulative energy percentage of each principal component in turn.

[0057] In a specific embodiment, suppose that the covariance matrix of a multimodal prediction data, after singular value decomposition, yields four principal components with eigenvalues ​​of 20, 8, 1.5, and 0.5. To analyze the contribution of each principal component to the overall variance, these four eigenvalues ​​are first sorted from largest to smallest (already sorted in this example). Then, the sum of all eigenvalues ​​is calculated: total energy = 20 + 8 + 1.5 + 0.5 = 30. The energy percentage and cumulative energy percentage of each principal component are calculated as follows: First principal component (eigenvalue 20): energy percentage is 20 / 30 × 100% = 66.7%, cumulative energy percentage is 66.7%. Second principal component (eigenvalue 8): energy percentage is 8 / 30 × 100% = 26.7%, cumulative energy percentage is (20 + 8) / 30 × 100% = 28 / 30 × 100% = 93.3%. The third principal component (eigenvalue 1.5): its energy percentage is 1.5 / 30×100%=5%, and its cumulative energy percentage is (20+8+1.5) / 30×100%=29.5 / 30×100%=98.3%. The fourth principal component (eigenvalue 0.5): its energy percentage is 0.5 / 30×100%=1.7%, and its cumulative energy percentage is 100%.

[0058] Setting a cumulative energy percentage threshold requires clarification. It's important to note that this threshold is typically set to explain the majority of the variance. Common practices include integer percentages such as 90%, 95%, 98%, and 99%. In practical data analysis, signal processing, PCA, and spectral correction, 95% is the most frequently used empirical value. It's worth noting that in actual implementation, the cumulative energy percentage threshold can be flexibly adjusted based on business needs. This example only provides a numerical illustration and does not impose specific limitations.

[0059] When the cumulative energy percentage of each principal component exceeds the cumulative energy percentage threshold for the first time, the corresponding number of principal components is obtained and recorded as the number of qualified principal components.

[0060] Continuing with the above embodiment, assuming the cumulative energy percentage threshold is 95%, and the cumulative energy percentage of the third principal component is 98.3%, exceeding 95% for the first time, then the number of qualified principal components in this embodiment is 3.

[0061] Obtain the total number of principal components, set the principal component validation percentage, and use the product of the total number of principal components and the principal component validation percentage as the principal component validation count.

[0062] In this embodiment, the percentage of principal component number verification is preferably set to 70%. In actual applications, the percentage of principal component number verification is usually between 50% and 80%, and the specific value can be flexibly adjusted according to actual system requirements and experience.

[0063] If the number of qualified principal components is greater than or equal to the number of principal component validations, it indicates that the spectral structure of the current covariance matrix is ​​uniformly distributed, the principal component energies are not excessively concentrated, and the correlation between variables is reasonable. At this time, the system does not exhibit significant low-rank or ill-conditioned phenomena, the sample set can fully characterize the main variance structure of the original space, and the diversity and richness of the data are guaranteed. Therefore, the covariance matrix is ​​directly used as the optimized covariance matrix.

[0064] If the number of qualified principal components is less than the number of principal component validations, it indicates that the covariance matrix exhibits excessive concentration of principal component energy, with most variance information distributed only in a few principal component directions. This reflects ill-conditioned characteristics such as low rank, redundancy, or insufficient sampling in the data. In this case, the sample set cannot effectively cover the multidimensional structure of the original space, which can easily lead to risks such as unstable filter values, ill-conditioned gain, abnormal amplification, or compression. Therefore, spectral correction is required, and the covariance matrix after spectral correction is used as the optimized covariance matrix.

[0065] Further, spectral correction is performed, the specific process of which is as follows:

[0066] The number of qualified principal components is obtained by subtracting the number of qualified principal components from the number of verified principal components.

[0067] The method for finding the proportion of additional retained principal components based on the number of qualified principal component deviations involves constructing a mapping relationship system between the number of qualified principal component deviations and the proportion of additional retained principal components. This system needs to be established by combining the characteristic attributes of the target space multimodal sensing data (including but not limited to the sampling frequency, spatial resolution, and noise level of modal data such as temperature, airflow, air quality, and personnel distribution), the system's preset filtering accuracy requirements (such as state estimation error thresholds and data fusion real-time indicators), and historical spectrum correction effect feedback data, through a combination of offline simulation and engineering field measurements. Specifically, the system first simulates a large amount of multimodal prediction data under various operating conditions (such as drastic changes in airflow in the target space, data loss due to temporary sensor failures, and spatiotemporal synchronization deviations in multimodal data) to calculate the optimal proportion of additional retained principal components corresponding to different numbers of qualified principal component deviations. This means that the covariance matrix after spectral correction at this proportion can effectively avoid filter divergence and ensure that there is no excessive loss or redundancy of information during the state update process. Then, the correspondence between the number of deviations and the proportion obtained from the above simulation is iteratively optimized in combination with the spectral correction effect under different deviation scenarios in the actual operation of the system (with the multimodal state vector output accuracy and filter convergence speed as evaluation indicators). Finally, a standardized mapping table covering the entire deviation range (such as 0-10 deviations) is formed. Each qualified principal component deviation in the mapping table corresponds to a unique proportion of additional retained principal components, and the proportion value must be limited to between 5% and 30% (balancing filtering accuracy and computational efficiency, avoiding data redundancy due to an excessively high proportion and information loss due to an excessively low proportion). Secondly, in the real-time spectral correction process of the system, after the number of qualified principal component deviations is calculated using the above method, the pre-constructed mapping table of qualified principal component deviations - additional retained principal component ratios is called to perform precise matching based on the currently calculated number of qualified principal component deviations. If the number of deviations is an integer and within the preset range of the mapping table, the corresponding additional retained principal component ratio is directly extracted; if the number of deviations is non-integer due to special operating conditions (such as a sudden change in the sampling frequency of multimodal data causing decimal deviations in the principal component analysis results) or exceeds the preset range of the mapping table, the interpolation completion mechanism is activated. For non-integer deviations, linear interpolation is used to calculate the current deviation ratio based on the ratio values ​​corresponding to the two adjacent integer deviations; for qualified principal component deviations exceeding the preset range, if the number of qualified principal component deviations is greater than the upper limit of the range, the ratio corresponding to the largest deviation in the mapping table is taken as the current ratio (to avoid a surge in computational load due to an excessively high ratio), and if the number of qualified principal component deviations is less than the lower limit of the range (i.e., close to 0), the ratio corresponding to the smallest deviation in the mapping table is taken (to ensure the retention of basic information).

[0068] The product of the total number of principal components and the proportion of additional principal components retained is taken as the number of additional principal components retained.

[0069] The number of retained principal components is obtained by subtracting the number of additional retained principal components from the number of qualified principal components.

[0070] Based on the descending order of the principal component eigenvalues ​​and the number of principal components to be retained, the principal components to be retained are determined, and the remaining principal components are truncated to obtain the truncated covariance matrix.

[0071] In one specific embodiment, assume that the covariance matrix, after principal component decomposition, yields 10 principal components. The system sets the principal component validation percentage to 70%, meaning the number of validated principal components is 10 × 0.7 = 7. The cumulative energy percentage threshold is set to 95%. Actual calculations show that only the first 5 principal components are needed to reach 95% cumulative energy, therefore the number of qualified principal components is 5. Based on the above conditions, the number of qualified principal component deviations is 7 - 5 = 2. Furthermore, the system sets the additional principal component retention ratio to 10%, so the number of additional retained principal components is 10 × 0.1 = 1. Finally, the system adds the number of qualified principal components to the number of additional retained principal components, resulting in a total of 6 principal components to be retained. Subsequently, based on the 10 principal components sorted by eigenvalue from largest to smallest, the first 6 principal components are retained, and the eigenvalues ​​of the remaining 4 principal components are truncated (e.g., set to zero or assigned a small positive number). The covariance matrix is ​​then reconstructed accordingly, forming the truncated optimized covariance matrix. This processing method not only ensures coverage of the energy accumulation threshold, but also reduces the risk of low-rank and ill-conditioned spectral structures by supplementing with additional principal components, providing a more robust numerical foundation for subsequent data fusion and filtering algorithms.

[0072] Numerical smoothing is performed on the truncated covariance matrix, and the smoothed truncated covariance matrix is ​​used as the optimized covariance matrix. Specifically, the truncated principal component eigenvalue sequence (i.e., the spectrum formed by retaining only some principal components and setting the remaining eigenvalues ​​to zero or assigning small positive numbers) is smoothed. Common methods include using moving averages or exponentially weighted averages for continuous principal component eigenvalues ​​to eliminate discontinuities, abrupt changes, or sharp drops in the eigenspectrum that may be caused by direct truncation. Secondly, the covariance matrix can be reconstructed based on the smoothed spectrum and the original eigenvector directions to ensure a smoother transition in the spectrum distribution and no significant numerical abrupt changes.

[0073] It should be explained that, in this embodiment, the main purpose of numerical smoothing is to avoid abrupt changes or discontinuities in the covariance spectrum, thereby reducing error amplification or loss of abnormal information caused by matrix singularity or ill-conditionedness during numerical optimization and filtering recursion; to improve the positive definiteness of the matrix, the coherence and physical rationality of the principal component energy distribution, and to ensure the numerical robustness and physical interpretation capability of subsequent data fusion, Kalman gain calculation and other links, thereby ensuring the stability of the entire system and the reliability of practical engineering.

[0074] like Figure 3The diagram shown is a flowchart of the optimized state-observation cross-covariance matrix analysis involved in this embodiment. First, the state-observation cross-covariance matrix is ​​calculated, and the historical average state-observation cross-covariance matrix is ​​analyzed. Then, the matrix distance between the current state-observation cross-covariance matrix and the historical average state-observation cross-covariance matrix is ​​calculated using the norm formula. Next, it is determined whether the matrix distance is greater than a matrix distance threshold: if the matrix distance is less than or equal to the matrix distance threshold, the state-observation cross-covariance matrix is ​​directly used as the optimized state-observation cross-covariance matrix; if the matrix distance is greater than the matrix distance threshold, cross-covariance coefficient reduction is performed, specifically including analyzing the matrix distance deviation value, determining the cross-covariance coefficient reduction value, and reducing the state-observation cross-covariance matrix based on this reduction value; finally, the optimized state-observation cross-covariance matrix is ​​obtained.

[0075] Furthermore, based on the optimized covariance matrix, the state-observation cross-covariance matrix is ​​calculated, and the optimized state-observation cross-covariance matrix is ​​obtained through analysis. The specific acquisition process is as follows:

[0076] Based on the optimized covariance matrix, the state-observation cross-covariance matrix is ​​calculated as follows: , where P xz Let H be the state-observation cross covariance matrix, P be the optimized covariance matrix, and H be the state-observation cross covariance matrix. T This is the transpose of the Jacobian matrix.

[0077] It should be noted that the state-observation cross-covariance matrix is ​​a matrix used to describe the correlation structure of the joint distribution between the system's state variables and observed variables. It quantitatively reflects the degree of linear correlation between the state space and the observation space.

[0078] In this embodiment, the transpose of the Jacobian matrix refers to the operation of interchanging the rows and columns of the first-order partial derivative matrix of the observation function with respect to the system state variables. For an observation function h(x) from an n-dimensional state space to an m-dimensional observation space, the elements of its Jacobian matrix H are: That is, the partial derivative of the j-th observation with respect to the k-th state variable, where x k Let h represent the k-th state variable (k is the state variable number, k=1,2,,,n). j Let represent the j-th observation (j is the observation number, j=1,2,,,m), therefore the Jacobian matrix is ​​an m×n matrix. Interchanging the rows and columns of this matrix yields the transpose H of the Jacobian matrix. T Its dimension is n×m. In practical engineering, if the form of the observation function is known, the partial derivatives of each output component with respect to all state variables can be directly taken to form the Jacobian matrix, and H can be obtained by the standard matrix transpose operation. TFor observation functions that are complex or lack analytical expression, numerical methods such as automatic differentiation and finite difference or mainstream mathematical software tools can be used to automatically obtain the Jacobian matrix, and then perform a transpose operation to ensure that subsequent filtering, fusion and other processes accurately model the correlation between the system state and the observation.

[0079] It should be added that the observation function describes the mapping relationship between system state variables and actual observations, serving as a mathematical bridge connecting the state space and the observation space. The specific form of the observation function is determined by the system's physical mechanism, sensor type, and actual measurement method. For different engineering applications, the observation function can be linear or highly nonlinear, and its expression can be obtained through theoretical modeling, or determined based on experimental data fitting, system identification, or by referring to sensor technical manuals. In Kalman filtering and its extended algorithms, the accuracy of the observation function directly affects the final results of state estimation and data fusion. Therefore, in the actual design and implementation process, the physical rationality, differentiability, and engineering availability of the observation function should be fully considered to ensure the accurate acquisition of the Jacobian matrix and its transpose, and the effective operation of subsequent filtering algorithms.

[0080] Obtain the set of historical state-observation cross-covariance matrices, and calculate their moving average to obtain the historical average state-observation cross-covariance matrix. That is, during multiple filtering cycles of continuous system operation, the state-observation cross-covariance matrix calculated based on the optimized covariance matrix and the Jacobian matrix of the observation model is recorded sequentially for each cycle. The cross-covariance matrices of the most recent N cycles are stored in chronological order to form the set of historical state-observation cross-covariance matrices.

[0081] It should be noted that in this embodiment, the value of N in the most recent N periods can be flexibly set according to the system's dynamic characteristics, observation frequency, and service response requirements. Generally, N is preferably between 10 and 30 to balance the smoothness of historical statistics and the responsiveness to the latest changes in the system's state. The specific value can be further adjusted based on the actual operating environment of the system and simulation verification results.

[0082] The matrix distance between the historical average state-observation cross-covariance matrix and the state-observation cross-covariance matrix is ​​calculated using the norm formula, specifically as follows: , Where d is the matrix distance, ΔP is the difference matrix, and P xz (t) Let the state-observation cross-covariance matrix be the state-observation cross-covariance matrix for the current period (t). This is the historical average state-observation cross-covariance matrix.

[0083] If the matrix distance exceeds a preset threshold, it indicates a significant deviation between the current period's state-observation cross-covariance matrix and the historical average state-observation cross-covariance matrix. This suggests an abnormal change in the correlation structure between the system state and observations, potentially caused by sudden disturbances, insufficient samples, model drift, or sensor malfunctions. In this case, to prevent the abnormal covariance structure from affecting the calculation of the Kalman gain and the numerical stability of the filtering system, the cross-covariance coefficients are reduced, thus obtaining the optimized state-observation cross-covariance matrix.

[0084] Furthermore, the cross-covariance coefficient is reduced to obtain the optimized state-observation cross-covariance matrix. The specific execution process is as follows:

[0085] Subtract the matrix distance threshold from the matrix distance to obtain the matrix distance deviation value.

[0086] It should be noted that the matrix distance threshold is a baseline parameter preset in the database during the initial deployment or operation of the system, based on historical data statistical analysis and expert experience. The matrix distance threshold reflects the maximum tolerable fluctuation range between the state-observation cross-covariance matrix and its historical average state-observation cross-covariance matrix under normal operating conditions. Its setting rules typically include: based on large-scale historical operating data of the system, statistically analyzing the distribution characteristics of the matrix distance over all periods (such as mean, standard deviation, quantiles, etc.), selecting a boundary value that can effectively distinguish between normal fluctuations and abnormal variations as the matrix distance threshold (e.g., taking the distribution mean plus twice the standard deviation, or using the 99th quantile, etc.); and appropriately adjusting it based on the actual business requirements for sensitivity and fault tolerance.

[0087] The method for determining the reduction value of the cross-covariance coefficient based on the matrix distance deviation value is as follows: A mapping system of matrix distance deviation value and cross-covariance coefficient reduction value is constructed. Specifically, firstly, the optimal reduction value of the cross-covariance coefficient corresponding to each deviation value is collected by simulating matrix distance deviation value scenarios under different operating conditions (such as sudden changes in airflow in the target space, temporary sensor failures, and multimodal data synchronization deviations). This reduction value ensures that the adjusted cross-covariance matrix eliminates the influence of abnormal deviations without losing effective state-observation correlation information, and guarantees the numerical stability of subsequent Kalman gain calculations. Then, the correspondence between the deviation values ​​and reduction values ​​obtained from the simulation is iteratively optimized by combining the filtering effect under different deviation value scenarios in actual system operation (using multimodal state vector output accuracy and filtering convergence speed as evaluation indicators). Finally, a standardized mapping table covering the entire deviation value range (e.g., 0-100) is formed. Each matrix distance deviation value in the mapping table corresponds to a unique cross-covariance coefficient reduction value, and the reduction value is limited to the range of 0.1-1 (to avoid excessive reduction leading to information distortion or insufficient reduction failing to correct anomalies). During real-time system operation, after calculating the matrix distance deviation value, the system first determines the interval level to which the matrix distance deviation value belongs, and then calls the corresponding interval's reduction value adjustment rule to match it with the mapping table. If the matrix distance deviation value is an integer and within the preset interval of the mapping table, the corresponding cross-covariance coefficient reduction value is directly extracted. If, due to special operating conditions (such as sudden changes in the sampling frequency of multimodal data), the deviation value becomes a non-integer or exceeds the preset interval of the mapping table, the interpolation completion and boundary constraint mechanism is activated. For non-integer deviation values, linear interpolation is used, and the reduction value of the current matrix distance deviation value is calculated based on the reduction values ​​corresponding to its two adjacent integer deviation values. When the matrix distance deviation value exceeds the upper limit of the interval, the reduction value corresponding to the largest deviation value in the mapping table is taken (to avoid excessive reduction leading to information loss). When the matrix distance deviation value is lower than the lower limit of the interval, the reduction value corresponding to the smallest deviation value in the mapping table is taken (to ensure the basic adjustment effect).

[0088] The state-observation cross-covariance matrix is ​​reduced based on the reduced cross-covariance coefficient value. Specifically, the reduced cross-covariance coefficient value for the current period is determined (usually a positive real coefficient less than 1). Then, all elements of the current state-observation cross-covariance matrix are multiplied by this reduced cross-covariance coefficient value to obtain the reduced cross-covariance matrix, i.e.: , where P xz (scaled) The reduced cross-covariance matrix (i.e., the optimized state-observation cross-covariance matrix), where α is the reduced value of the cross-covariance coefficients, and P... xz (t) Let be the state-observation cross-covariance matrix for the current period (t).

[0089] The reduced state-observation cross-covariance matrix is ​​used as the optimized state-observation cross-covariance matrix.

[0090] If the matrix distance is less than or equal to the matrix distance threshold, then the state-observation cross-covariance matrix is ​​directly used as the optimized state-observation cross-covariance matrix.

[0091] The Kalman gain matrix is ​​calculated based on the optimized covariance matrix and the optimized state-observation cross covariance matrix, and the optimized Kalman gain matrix is ​​obtained through analysis.

[0092] The Kalman gain matrix is ​​calculated according to the following standard formula: , Where K is the Kalman gain matrix, P xz (scaled) To optimize the state-observation cross-covariance matrix, let S be the observation-prediction covariance matrix, (S) -1 H is the inverse of the observed prediction covariance matrix, and H is the Jacobian matrix. T Let P be the transpose of the Jacobian matrix, P be the optimized covariance matrix, and R be the observation noise covariance matrix.

[0093] It should be noted that the observation noise covariance matrix is ​​used to quantify the variance and correlation of noise in each measurement channel of the observation system. It can be obtained through statistical analysis of measured data, sensor technical documentation, theoretical estimation, or online adaptive methods.

[0094] Furthermore, the optimized Kalman gain matrix was obtained through analysis. The specific analysis process is as follows:

[0095] The Kalman gain norm is calculated using the norm formula based on the Kalman gain matrix.

[0096] Query the historical Kalman gain matrix of the sliding window records, calculate the historical average Kalman matrix, and calculate the historical Kalman gain norm using the norm formula.

[0097] To find the Kalman gain norm verification scaling factor set based on the reduced cross-covariance coefficient value, the following steps are taken: A mapping system is constructed between the reduced cross-covariance coefficient value and the Kalman gain norm verification scaling factor set. Specifically, firstly, the optimal Kalman gain norm verification scaling factor set corresponding to each reduced value is collected by simulating different scenarios of reduced cross-covariance coefficient values ​​under various operating conditions (such as airflow disturbance in the target space, sensor accuracy drift, and multimodal data synchronization deviation). This set must satisfy the following: based on the Kalman gain norm allowable interval constructed therein, it can effectively filter out abnormal gain matrices and adapt to the state-observation correlation strength under the current reduced value, avoiding excessive constraints leading to conservative filtering or insufficient constraints causing numerical divergence. Then, the correspondence between the reduced value obtained from the simulation and the scaling factor set is iteratively optimized by combining the filtering effect under different reduced value scenarios in actual system operation (using multimodal state vector output accuracy and filtering iteration convergence as evaluation indicators), ultimately forming a standardized mapping table covering the entire reduced value interval. In the mapping table, each reduced cross-covariance coefficient value corresponds to a unique set of Kalman gain norm verification scaling factors. This set includes both a first and a second Kalman gain norm verification scaling factor, ensuring the compatibility between the scaling factors and the reduced value. During real-time system operation, once the reduced cross-covariance coefficient value is calculated, it is used as the query key to retrieve the corresponding Kalman gain norm verification scaling factor set within the mapping system of reduced cross-covariance coefficient value and Kalman gain norm verification scaling factor set.

[0098] The set of Kalman gain norm verification scaling factors includes a first Kalman gain norm verification scaling factor (greater than zero and less than 1) and a second Kalman gain norm verification scaling factor (greater than 1).

[0099] The product of the historical Kalman norm and the verification scaling factor of the first Kalman gain norm is used as the upper limit of the allowable interval of the Kalman gain norm, and the product of the historical Kalman norm and the verification scaling factor of the second Kalman gain norm is used as the lower limit of the allowable interval of the Kalman gain norm, thus obtaining the allowable interval of the Kalman gain norm.

[0100] If the Kalman gain norm falls within the allowable range, it indicates that the current state-observation correlation structure matches the historical distribution, and the system observation and state estimation processes are in a numerically stable and physically reasonable normal operating range. In this case, no special correction is needed for the Kalman gain; the current period's Kalman gain matrix can be directly used as the optimized Kalman gain matrix for state vector updates and recursive estimation, ensuring the efficiency and response sensitivity of the filtering system.

[0101] If the Kalman gain norm does not fall within the allowable range, it indicates an abnormal fluctuation in the current observation or state estimation process. This may be caused by factors such as sample anomalies, noise anomalies, observation model mismatch, or extreme perturbations, resulting in an excessively large or small Kalman gain value. To prevent errors from being amplified or useful information from being suppressed due to abnormal gain, the system will automatically perform scaling of the Kalman gain matrix and use the scaled Kalman gain matrix as the optimized Kalman gain matrix.

[0102] Furthermore, scaling of the Kalman gain matrix is ​​performed, and the specific execution process is as follows:

[0103] If the Kalman gain norm is less than the lower limit of the allowable range of the Kalman gain norm, then the first Kalman gain bias norm is obtained by subtracting the Kalman gain norm from the lower limit of the allowable range. The Kalman gain bias-Kalman gain matrix reduction factor configuration table is then extracted. Based on the first Kalman gain bias norm, the Kalman gain matrix reduction factor is queried. The specific query method is as follows: First, the system-pre-stored Kalman gain bias-Kalman gain matrix amplification factor configuration table is called. This table is obtained by offline simulation of different first Kalman gain bias norm scenarios, combined with measured filtering effects (based on state estimation accuracy). The Kalman gain deviation norm is calibrated using convergence speed as an indicator. Each deviation norm interval or specific value in the table corresponds to a unique reduction factor (the larger the deviation, the smaller the factor, to avoid excessive reduction). The calculated first Kalman gain deviation norm is then matched with the data in the table. If the first Kalman gain deviation norm is a specific value in the table, the corresponding reduction factor is directly extracted. If it is a value within an interval, linear interpolation is used for calculation. If it exceeds the maximum deviation norm in the table, the reduction factor corresponding to the maximum deviation in the table is taken to ensure that the query results are adapted to the current gain deviation state and to ensure the rationality of subsequent Kalman gain matrix scaling and filtering stability.

[0104] If the Kalman gain norm is greater than the upper limit of the allowable range of the Kalman gain norm, then the second Kalman gain bias norm is obtained by subtracting the upper limit of the allowable range from the Kalman gain norm. The Kalman gain bias-Kalman gain matrix amplification coefficient configuration table is then extracted. Based on the second Kalman gain bias norm, the Kalman gain matrix amplification coefficient is queried. The specific query method is as follows: First, the Kalman gain bias-Kalman gain matrix amplification coefficient configuration table pre-stored in the system is called. This table is obtained by offline simulation of different second Kalman gain bias norm scenarios, combined with measured filtering effects (based on state estimation accuracy, numerical...). The value stability is used as an indicator for calibration. Each deviation norm interval or specific value in the table corresponds to a unique amplification factor (the smaller the deviation, the closer the factor is to 1, to avoid over-amplification). Then, the calculated second Kalman gain deviation norm is matched with the data in the table. If the second Kalman gain deviation norm is a specific value in the table, the corresponding amplification factor is directly extracted. If it is a value within the interval, linear interpolation is used for calculation. If it exceeds the maximum deviation norm in the table, the amplification factor corresponding to the maximum deviation in the table is taken to ensure that the query result is adapted to the current gain deviation state and to ensure the rationality of subsequent Kalman gain matrix scaling and filtering convergence.

[0105] Scaling the Kalman gain matrix is ​​performed based on either the scaling factor or the scaling factor, which means multiplying all elements of the Kalman gain matrix by the scaling factor or the scaling factor to obtain the optimized Kalman gain matrix.

[0106] Based on an optimized Kalman gain matrix, this system achieves the fusion and recursion of multimodal sensing and prediction data to obtain a multimodal state vector. Specifically, the multimodal sensing data is transformed into a multimodal sensing observation vector, and the multimodal prediction data is transformed into a multimodal prediction state vector. The multimodal prediction state vector is mapped to the observation space using an observation function to obtain multimodal prediction observations. The observation residuals are calculated by subtracting the multimodal prediction observations from the multimodal sensing observation vectors. Using the optimized Kalman gain matrix, the observation residuals are weighted and converted into corrections to the predicted state, which are then added to the original predicted state vector to form the updated multimodal state vector for the current cycle. This updated state vector not only reflects the optimal fusion result of model prediction and actual observation but also possesses dynamic recursive capabilities, serving as the basis for prediction and fusion in the next cycle. The entire process achieves effective fusion and recursive optimization of multimodal prediction and sensing data, ensuring high-precision estimation and real-time updates of states in complex environments, providing a reliable data foundation for subsequent risk assessment and intelligent decision-making modules.

[0107] The aerosol risk prediction and intelligent decision-making module is used to dynamically predict the airflow and aerosol diffusion process in the target space based on multimodal state vectors and computational fluid dynamics models, construct a risk assessment model, and generate corresponding thermal barrier deployment decision instructions.

[0108] Furthermore, a risk assessment model is constructed, and corresponding thermal barrier deployment decision instructions are generated. The specific execution process is as follows:

[0109] The process of resolving physical field parameter subsets based on multimodal state vector analysis typically includes the following steps: First, based on the physical field modeling or simulation requirements, determine the key physical quantities of interest in the target space, such as temperature, airflow velocity, pressure, humidity, and CO2 concentration. After filtering and fusion, the multimodal state vector contains the optimal estimates of various observational and model prediction data. The system then extracts elements corresponding to the required physical field parameters from the multimodal state vector based on parameter definitions, variable indices, or labels, forming a subset of physical field parameters. For example, temperature field-related components, airflow vector components, and gas concentration components can be selected in batches through variable name mapping, structured data indexing, or label matching, while data irrelevant to this stage of modeling can be removed.

[0110] A computational fluid dynamics model is established, and a subset of physical field parameters is used as input to predict the time series of airflow and aerosol diffusion paths in the target space, outputting the space airflow velocity field and aerosol concentration distribution for future time periods.

[0111] In a specific embodiment, establishing a computational fluid dynamics (CFD) model typically includes the following steps: First, based on the geometry of the target space, import or construct a Building Information Modeling (BIM), a 2D / 3D CAD model, or directly use spatial data collected by methods such as LiDAR or point clouds to determine the physical domain and mesh generation for fluid simulation. Second, combine the subset of physical field parameters obtained from multimodal state vector analysis (such as temperature distribution, airflow velocity, humidity, CO2 concentration, etc.) to provide initial and boundary conditions for the CFD model, including inlet velocity, temperature, concentration, wall conditions, source term distribution, etc. Then, select appropriate fluid dynamics equations (such as incompressible Navier-Stokes equations, energy equations, species diffusion equations, etc.) and physical modules such as turbulence models and heat transfer models to model airflow, temperature, and aerosol diffusion under complex environments.

[0112] It is important to note that the CFD model does not limit the input data to only a few physical quantities such as temperature, airflow velocity, pressure, humidity, and CO2 concentration. These parameters are merely illustrative examples of commonly used physical quantities in the current embodiment. In specific applications, the system's input data can also include, but is not limited to, particulate matter (PM2.5, PM10) concentration, formaldehyde / volatile organic compound (VOC) concentration, light intensity, noise level, surface temperature, infrared / ultraviolet radiation intensity, air negative ion concentration, and even the operating status of specific equipment, thermal estimation of personnel activity, and spatial pressure differences. In other words, the type and quantity of input data can be flexibly expanded and are not limited by the aforementioned exemplified parameters. It can be dynamically adjusted and optimized according to the actual application scenario and data availability to adapt to various complex and changing spatial environments and engineering needs.

[0113] The system, based on professional CFD software (such as ANSYS Fluent, OpenFOAM, COMSOL, etc.) or a customized numerical solver, automatically establishes simulation cases, configures physical parameters, and sets the solution step size and convergence conditions. Through solving, it obtains the airflow velocity field, temperature field, and pollutant concentration distribution in the target space at different times, achieving high-precision simulation of the dynamic evolution of the physical field. This process can be embedded into the system's automated management, supporting real-time simulation, parameter iteration, and multi-objective optimization, ensuring the scientific rigor and intelligence of thermal barrier deployment and environmental control.

[0114] An assessment model is constructed, which takes the spatial airflow velocity field and aerosol concentration distribution in future time periods as inputs, outputs a comprehensive risk index, and determines the risk level.

[0115] In a specific embodiment, an assessment model is established, and the specific steps include: First, spatial discretization and feature extraction are performed on the airflow velocity field and aerosol concentration distribution output by CFD to obtain representative parameters (such as maximum concentration, average concentration, percentage of exceeding standards, ventilation dead zones, etc.) for each key location or region. Second, a multi-index weighted or multi-dimensional normalized risk assessment function is constructed, and the parameters for each time point and each region are standardized and scored based on historical statistical experience. Through expert experience, machine learning algorithms, or multi-criteria decision-making methods (such as AHP, fuzzy comprehensive evaluation, etc.), the various sub-indicators are integrated into a comprehensive risk index. Third, based on the comprehensive risk index, grading rules or segmented thresholds are set to automatically determine the risk level (such as low risk, medium risk, high risk, extremely high risk), and output a visualized risk heat map or early warning results. The entire assessment process can achieve parameterized configuration, automated execution, and real-time updates, supports multi-scenario switching and adaptive optimization, and provides data support and risk early warning basis for thermal barrier deployment decisions and environmental safety control.

[0116] Based on the risk level, the system outputs dynamic deployment parameters, including thermal barrier height, spacing, angle, and transmittance, thereby generating thermal barrier deployment decision instructions. The specific steps are as follows: First, the system establishes a mapping rule base from parameters to equipment control instructions, automatically matching appropriate thermal barrier unit configuration schemes according to different risk levels and spatial distributions. For example, for high-risk areas, the thermal barrier height can be set higher, the spacing reduced, and the angle perpendicular to the main airflow direction, while increasing the shading rate of the thermal barrier units or decreasing the transmittance. Second, the deployment parameters are formatted and converted according to equipment communication protocols and execution standards to generate standardized control instructions including thermal barrier type, unit number, spatial coordinates, geometric attitude (height, angle), and functional parameters (transmittance, on / off status). The system can automatically issue instructions to field robots, thermal barrier actuators, or intelligent gateways through the task scheduling module, realizing the automatic deployment, adjustment, and removal of thermal barrier units.

[0117] The intelligent thermal barrier generation and automatic deployment module is used to control the parameter execution and spatial arrangement of adjustable thermal barrier units with shape memory alloy skeleton and aerogel composite structure according to thermal barrier deployment decision instructions, and drive the robot cluster to automatically complete the deployment, adjustment and removal of thermal barriers to achieve spatial thermal isolation and environmental control.

[0118] The shape memory alloy skeleton has a phase transformation temperature range of 30-60 degrees Celsius, an aerogel layer thickness of 5-20 mm, and an adjustable light transmittance range of 10%-80%.

[0119] Specifically, the system first parses the thermal barrier deployment decision commands, extracting the spatial arrangement parameters and functional settings of each thermal barrier unit, including target location, installation height, spacing, angle, light transmittance, and opening / closing status. Subsequently, the system assigns these parameters to adjustable thermal barrier units with shape memory alloy skeletons and aerogel composite structures, and generates specific posture adjustment and function switching commands based on their structural characteristics. After receiving the assigned control commands, the robot swarm automatically proceeds to the designated deployment location using autonomous navigation, obstacle avoidance, and cooperative positioning technologies. During the on-site operation phase, the robots utilize high-precision positioning and actuators to achieve precise placement, posture adjustment, and functional activation of the thermal barrier units. For example, through electric heating and stress release, the shape memory alloy skeleton is driven to unfold the thermal barrier unit to a specified height and angle, while simultaneously adjusting the arrangement of the aerogel layers to set the light transmittance and thermal insulation performance. For thermal barrier units that need adjustment or retraction, the robots automatically complete folding, storage, and relocation according to the latest commands, achieving flexible isolation and real-time optimization and control of the spatial thermal environment. Throughout the process, the robot cluster can achieve dynamic task allocation and multi-machine collaboration, ensuring efficient, safe and intelligent response in thermal barrier deployment.

[0120] In this embodiment, the deployment stage employs a thermal barrier deployment robot cluster composed of an AGV chassis and a six-degree-of-freedom robotic arm. It supports magnetic module connection, has a maximum deployment speed of ≥1.0 m / s, and a single charge provides ≥4 hours of battery life. Deployment commands are derived from real-time risk assessment and environmental perception data. The robot can autonomously plan its path, automatically connecting thermal barrier modules and adjusting their spacing and height, achieving rapid erection and removal of spatial thermal barriers within minutes.

[0121] The digital twin verification and feedback optimization module is used to build a digital twin scenario based on building information model and real-time multimodal perception data in a virtual simulation platform, to conduct simulation verification and effect evaluation of thermal barrier deployment decisions, and to feed back the verification results to the multimodal data perception and fusion module and the aerosol risk prediction and intelligent decision-making module to realize the automatic closed-loop optimization of the system.

[0122] Furthermore, a digital twin scenario is established in the virtual simulation platform based on the building information model and real-time multimodal sensing data to conduct simulation verification and effect evaluation of thermal barrier deployment decisions. The specific execution process is as follows:

[0123] The building information model of the target space is imported into the virtual simulation platform, and the temperature field, airflow field and CO2 concentration distribution of the virtual space are dynamically updated based on real-time multimodal sensing data.

[0124] Based on the thermal barrier deployment scheme output by the aerosol risk prediction and intelligent decision-making module, the corresponding thermal barrier layout simulation is performed in a virtual scene to generate a comparison dataset before and after deployment.

[0125] The changes in airflow velocity, temperature field distribution, and air quality indicators in the virtual scene before and after deployment are calculated to obtain simulation evaluation results of the thermal barrier deployment effect.

[0126] Based on the simulation evaluation results of the thermal barrier deployment effect and the preset performance target, the dynamic deployment parameters are dynamically adjusted through optimization algorithms until the set thermal environment and air quality improvement targets are achieved.

[0127] The optimal parameters verified by simulation are fed back to the intelligent thermal barrier generation and automatic deployment module to guide the actual deployment and control of thermal barriers in physical space.

[0128] In one specific embodiment, the fever clinic waiting area has an area of ​​approximately 200 m² and a net height of 3 m, normally accommodating 40-60 patients simultaneously. The area contains various heat sources (human bodies, electronic devices, lighting) and airflow disturbances. The system installs the following equipment on the ceiling, walls, and floor: 4 infrared thermal imagers; 6 carbon dioxide sensors; 4 PM2.5 sensors; 3 ultrasonic anemometers and 2 wind vanes; 2 4K visible light cameras and 2 millimeter-wave radars; and 1 environmental monitoring module (for temperature, humidity, and air pressure data acquisition). The data sampling period is set to 2 seconds, and all sensors are synchronized via an edge gateway.

[0129] The collected multi-source data is input and an extended Kalman filter algorithm is executed to automatically fuse all real-time data, remove anomalies and noise, and generate the average temperature, carbon dioxide concentration, PM2.5 concentration, wind speed, and personnel density for each 4-square-meter grid in the waiting area. The system is set with the following automatic warning thresholds: if the temperature of any grid rises by more than 3 degrees Celsius within 10 minutes; the carbon dioxide concentration exceeds 1500 ppm; or the PM2.5 exceeds 75 μg / m³; if any of these indicators are triggered, the corresponding grid is judged as a high-risk area, and the system enters the intervention process.

[0130] The current status data is synchronously transmitted to the integrated CFD simulation module, automatically calling the BIM model of the waiting area and meshing it using 0.25-meter units. The system sets the air outlet velocity to 0.4 m / s, the wind vane to north, and boundary conditions to be measured in real time. The simulation duration is set to 15 minutes, outputting the temperature field, airflow distribution, carbon dioxide concentration, and PM2.5 diffusion path for the next 15 minutes. The simulation results automatically generate a risk zone heat map, marking high, medium, and low risk grids, and providing the optimal thermal barrier deployment scheme.

[0131] The central system issues control commands to dispatch three AGV thermal barrier deployment robots: each robot can carry two 1.2×2 meter shape memory alloy skeletons, 10mm thick, with a factory-set light transmittance of 60%; the robots deploy the thermal barrier modules to the boundaries of areas identified as high-risk according to the shortest path planning, with the deployment height automatically increasing to 2 meters, and the angle and light transmittance fine-tuned according to simulation suggestions (e.g., tilting the angle by 15 degrees and reducing the light transmittance to 30% to enhance the barrier effect). All deployment actions are completed within 3 minutes, during which the robots can temporarily change the deployment order and position based on real-time data.

[0132] All sensing and deployment data are synchronized in real time to the hospital's digital twin platform, dynamically displaying the current temperature field, airflow field, and carbon dioxide concentration distribution. The system automatically compares the changes in air quality indicators before and after deployment. If the temperature decrease is less than 30% or the carbon dioxide concentration decrease is less than 40%, a self-optimization process is automatically triggered, adjusting the height and angle of the thermal barrier modules or increasing the number of modules until the optimization standard is met. The optimization results are archived as parameters for subsequent system self-learning and rapid response.

[0133] The temperature in the high-risk grid of the waiting area dropped by 4.1 degrees Celsius, the carbon dioxide concentration decreased from 1650 ppm to 830 ppm, PM2.5 decreased to 35 μg / m³, and the airflow distribution became more uniform. The overall system energy consumption decreased by approximately 14% year-on-year, the response was fully automated, and the error between the digital twin platform and the on-site measurement data was less than 5%.

[0134] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0135] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0136] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0137] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0138] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0139] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0140] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0141] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0142] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A closed-loop intelligent thermal barrier system based on multimodal sensing and digital twin verification, characterized in that, include: The multimodal data perception and fusion module is used to collect multimodal perception data in the target space, perform spatiotemporal synchronization and coordinate alignment of the multimodal perception data, generate multimodal prediction data based on the state prediction model, and realize the fusion and recursion of multimodal perception data and multimodal prediction data through extended Kalman filtering, thereby outputting a multimodal state vector. The aerosol risk prediction and intelligent decision-making module is used to dynamically predict the airflow and aerosol diffusion process in the target space based on multimodal state vectors and computational fluid dynamics models, construct a risk assessment model, and generate corresponding thermal barrier deployment decision instructions. The intelligent thermal barrier generation and automatic deployment module is used to control the parameter execution and spatial arrangement of adjustable thermal barrier units with shape memory alloy skeleton and aerogel composite structure according to the thermal barrier deployment decision instructions, and drive the robot cluster to automatically complete the deployment, adjustment and removal of thermal barriers to achieve spatial thermal isolation and environmental control. The digital twin verification and feedback optimization module is used to establish a digital twin scenario based on building information model and real-time multimodal perception data in a virtual simulation platform, to conduct simulation verification and effect evaluation of thermal barrier deployment decisions, and to feed back the verification results to the multimodal data perception and fusion module and the aerosol risk prediction and intelligent decision-making module to realize the automatic closed-loop optimization of the system.

2. The closed-loop intelligent thermal barrier system based on multimodal perception and digital twin verification according to claim 1, characterized in that, The process of fusing and recursively applying extended Kalman filtering to multimodal sensing data and multimodal prediction data is as follows: The covariance matrix is ​​calculated based on multimodal prediction data, and the optimized covariance matrix is ​​obtained through analysis. Based on the optimized covariance matrix, the state-observation cross covariance matrix is ​​calculated, and the optimized state-observation cross covariance matrix is ​​obtained through analysis. The Kalman gain matrix is ​​calculated based on the optimized covariance matrix and the optimized state-observation cross covariance matrix, and the optimized Kalman gain matrix is ​​obtained through analysis. Based on the optimized Kalman gain matrix, the fusion and recursion of multimodal sensing data and multimodal prediction data are realized to obtain the multimodal state vector.

3. The closed-loop intelligent thermal barrier system based on multimodal perception and digital twin verification according to claim 2, characterized in that, The covariance matrix is ​​calculated based on multimodal prediction data, and the optimized covariance matrix is ​​obtained through analysis. The specific analysis process is as follows: Singular value decomposition is performed on the covariance matrix to obtain the eigenvalues ​​of each principal component; Sort the eigenvalues ​​of each principal component from largest to smallest, and calculate the cumulative energy percentage of each principal component in turn; Set a cumulative energy percentage threshold; When the cumulative energy percentage of each principal component exceeds the cumulative energy percentage threshold for the first time, the corresponding number of principal components is obtained and recorded as the number of qualified principal components. Obtain the total number of principal components, set the validation percentage for the principal components, and then obtain the number of validated principal components. If the number of qualified principal components is greater than or equal to the number of principal component verifications, then the covariance matrix is ​​directly used as the optimized covariance matrix. If the number of qualified principal components is less than the number of principal component verifications, then spectral correction is performed, and the covariance matrix after spectral correction is used as the optimized covariance matrix.

4. The closed-loop intelligent thermal barrier system based on multimodal perception and digital twin verification according to claim 3, characterized in that, The execution spectrum correction process is as follows: The number of qualified principal components is obtained by subtracting the number of verified principal components from the number of qualified principal components. The proportion of additional principal components to be retained is determined based on the number of qualified principal component deviations. The number of additional principal components is calculated based on the total number of principal components and the proportion of additional principal components retained. The number of retained principal components is calculated based on the number of qualified principal components and the number of additional retained principal components; Based on the descending order of the principal component eigenvalues, and combined with the number of principal components to be retained, the principal components to be retained are determined, and the remaining principal components are truncated to obtain the truncated covariance matrix. Numerical smoothing is performed on the truncated covariance matrix, and the numerically smoothed truncated covariance matrix is ​​used as the optimized covariance matrix.

5. The closed-loop intelligent thermal barrier system based on multimodal perception and digital twin verification according to claim 2, characterized in that, The optimized state-observation cross-covariance matrix is ​​calculated based on the optimized covariance matrix, and the optimized state-observation cross-covariance matrix is ​​obtained through analysis. The specific acquisition process is as follows: Based on the optimized covariance matrix, the state-observation cross covariance matrix is ​​calculated; Obtain the set of historical state-observation cross covariance matrices, and apply a moving average to them to obtain the historical average state-observation cross covariance matrix; The matrix distance between the historical average state-observation cross covariance matrix and the state-observation cross covariance matrix is ​​calculated using the norm formula. If the matrix distance is greater than the preset matrix distance threshold, then the cross covariance coefficient reduction is performed, thereby obtaining the optimized state-observation cross covariance matrix; If the matrix distance is less than or equal to the matrix distance threshold, then the state-observation cross-covariance matrix is ​​directly used as the optimized state-observation cross-covariance matrix.

6. The closed-loop intelligent thermal barrier system based on multimodal perception and digital twin verification according to claim 5, characterized in that, The process of reducing the cross-covariance coefficient is performed to obtain the optimized state-observation cross-covariance matrix. The specific execution process is as follows: The matrix distance deviation value is obtained by subtracting the matrix distance from the matrix distance threshold. Determine the reduction value of the cross covariance coefficient based on the matrix distance deviation value; Execution status based on reduced cross-covariance coefficient values ​​- observation of cross-covariance matrix reduction; The reduced state-observation cross-covariance matrix is ​​used as the optimized state-observation cross-covariance matrix.

7. The closed-loop intelligent thermal barrier system based on multimodal perception and digital twin verification according to claim 2, characterized in that, The analysis yields the optimized Kalman gain matrix, and the specific analysis process is as follows: The Kalman gain norm is calculated using the norm formula based on the Kalman gain matrix. Query the historical Kalman gain matrix of the sliding window records, calculate the historical average Kalman matrix, and calculate the historical Kalman gain norm using the norm formula; Find the Kalman gain norm verification scale factor set based on the reduced value of the cross-covariance coefficient; The set of Kalman gain norm verification scaling factors includes a first Kalman gain norm verification scaling factor and a second Kalman gain norm verification scaling factor; Based on the historical Kalman norm and Kalman gain norm verification scale factor set analysis, the allowable interval of the Kalman gain norm is obtained; If the Kalman gain norm falls within the allowable range of the Kalman gain norm, then the Kalman gain matrix is ​​directly used as the optimized Kalman gain matrix. If the Kalman gain norm does not fall within the allowed range of the Kalman gain norm, then the Kalman gain matrix is ​​scaled, and the scaled Kalman gain matrix is ​​used as the optimized Kalman gain matrix.

8. The closed-loop intelligent thermal barrier system based on multimodal perception and digital twin verification according to claim 7, characterized in that, The scaling of the Kalman gain matrix is ​​performed as follows: If the Kalman gain norm is less than the lower limit of the allowable range of the Kalman gain norm, the difference between the lower limit of the allowable range of the Kalman gain norm and the Kalman gain norm is processed to obtain the first Kalman gain deviation norm. The Kalman gain deviation-Kalman gain matrix reduction coefficient configuration table is extracted, and the Kalman gain matrix reduction coefficient is queried based on the first Kalman gain deviation norm. If the Kalman gain norm is greater than the upper limit of the allowable range of the Kalman gain norm, then the difference between the Kalman gain norm and the upper limit of the allowable range of the Kalman gain norm is processed to obtain the second Kalman gain deviation norm. The Kalman gain deviation-Kalman gain matrix amplification coefficient configuration table is extracted, and the Kalman gain matrix amplification coefficient is queried based on the second Kalman gain deviation norm. Scaling of the Kalman gain matrix is ​​performed based on either the scaling factor or the scaling factor of the Kalman gain matrix.

9. The closed-loop intelligent thermal barrier system based on multimodal perception and digital twin verification according to claim 1, characterized in that, The process of constructing a risk assessment model and generating corresponding thermal barrier deployment decision instructions is as follows: Based on the analysis of physical field parameter subsets using multimodal state vectors; A computational fluid dynamics model is established, and a subset of physical field parameters is used as input to predict the time series of airflow and aerosol diffusion paths in the target space, outputting the space airflow velocity field and aerosol concentration distribution for future time periods. An assessment model is constructed, taking the spatial airflow velocity field and aerosol concentration distribution in future time periods as inputs, and outputting a comprehensive risk index to determine the risk level. Based on the risk level, dynamic deployment parameters, including thermal barrier height, spacing, angle, and light transmittance, are output, thereby generating thermal barrier deployment decision instructions.

10. The closed-loop intelligent thermal barrier system based on multimodal perception and digital twin verification according to claim 1, characterized in that, The process of establishing a digital twin scenario based on building information modeling and real-time multimodal sensing data in a virtual simulation platform for simulation verification and effect evaluation of thermal barrier deployment decisions is as follows: The building information model of the target space is imported into the virtual simulation platform, and the temperature field, airflow field and CO2 concentration distribution of the virtual space are dynamically updated based on real-time multimodal sensing data. Based on the thermal barrier deployment scheme output by the aerosol risk prediction and intelligent decision-making module, the corresponding thermal barrier layout simulation is performed in the virtual scene to generate a comparison dataset before and after deployment. The changes in airflow velocity, temperature field distribution, and air quality indicators in the virtual scene before and after deployment are calculated to obtain simulation evaluation results of the thermal barrier deployment effect. Based on the simulation evaluation results of the thermal barrier deployment effect and the preset performance target, the dynamic deployment parameters are dynamically adjusted through optimization algorithm iteration until the set thermal environment and air quality improvement targets are achieved. The optimal parameters verified by simulation are fed back to the intelligent thermal barrier generation and automatic deployment module to guide the actual deployment and control of thermal barriers in physical space.

Images