BIM-based green smart hospital negative pressure ward dynamic airflow regulation method and system

Abstract

This invention discloses a BIM-based dynamic airflow control method and system for negative pressure wards in green and intelligent hospitals, belonging to the field of building intelligence and medical environment control technology. The method includes: constructing a BIM digital twin model and linking it with IoT data; deploying a multi-source sensor network to achieve real-time data fusion and quality control; defining ECA dynamic control rules and multi-objective optimization functions; achieving second-level CFD simulation and intelligent decision-making based on a reduced-order model; executing optimal control commands and forming closed-loop correction based on feedback; and realizing three-dimensional visualization monitoring, early warning, and full lifecycle data analysis. This invention, through a closed-loop intelligent control system of "perception-simulation-decision-execution-evaluation," transforms the airflow organization of negative pressure wards from static control to dynamic adaptation, and from experience-based control to predictive optimization decision-making. It optimizes system energy efficiency while ensuring infection control safety, significantly improving the level of intelligent operation and maintenance management.

Description

Technical Field

[0001] This invention belongs to the field of building intelligence and medical environment control technology, specifically involving a method for real-time, precise, and intelligent dynamic control of airflow organization in hospital negative pressure wards by combining building information modeling, Internet of Things sensing, artificial intelligence and computational fluid dynamics simulation. Background Technology

[0002] Negative pressure isolation wards are core medical facilities for treating patients with respiratory infectious diseases and preventing the spread of pathogens that could lead to nosocomial cross-infection. Their basic principle is to use a mechanical ventilation system to create a lower air pressure inside the ward compared to adjacent areas, forming a directional and stable pressure gradient from clean to semi-contaminated to contaminated zones, thus ensuring that contaminated air does not leak out. However, current mainstream negative pressure isolation ward airflow control technologies generally have the following limitations: (1) The control strategy is static and difficult to adapt to dynamic disturbances. Most existing systems are designed based on fixed operating conditions and adopt preset supply and exhaust air volumes and vent layouts. In actual operation, personnel entry and exit, equipment start and stop, door and window opening and closing, patient activities, etc. will dynamically change the indoor airflow field and pressure distribution. Static control strategies are difficult to continuously ensure the stability of pressure gradient and effective pollutant content under various disturbances.

[0003] (2) Reliance on manual intervention, resulting in delayed and inaccurate responses. The monitoring and control of key parameters such as pressure and differential pressure often rely on the experience and judgment of medical staff or facility maintenance personnel, lacking a real-time, automatic closed-loop control mechanism. In emergencies such as accidental door opening or equipment failure, the slow manual response poses a high risk of infection exposure.

[0004] (3) Lack of spatial visualization and predictive control capabilities. Maintenance personnel have difficulty intuitively and comprehensively grasping the airflow organization, pressure distribution, and potential diffusion paths of pollutants in the three-dimensional space inside the ward. Due to the lack of simulation and deduction tools based on real-time status, it is impossible to assess the effects of control actions before they are implemented, resulting in decision-making remaining in a passive mode of "experience-driven and post-event remediation".

[0005] Building Information Modeling (BIM) technology provides a high-fidelity digital carrier for building entities and functional attributes, but current applications are mostly concentrated in the design and construction phases, and there are still shortcomings in the operation and maintenance phase, especially in the deep integration with dynamic processes of medical infection control. Internet of Things (IoT) technology can realize the real-time collection of environmental data, but it is usually independent of the BIM model, and the data has a weak connection with spatial entities, making it difficult to support intelligent decision-making based on the coupling of spatial topology and real-time status.

[0006] To address the problems existing in the prior art, the present invention aims to provide a BIM-based dynamic airflow control method for negative pressure wards in green and intelligent hospitals. This method aims to solve the problems of static, lagging, and lacking visualization and predictive capabilities in traditional control methods. By constructing a closed-loop intelligent control system of "perception-simulation-decision-execution-evaluation," it achieves real-time, adaptive, and precise control of airflow organization in negative pressure wards. Under the premise of ensuring absolute safety in infection control, it optimizes system energy efficiency and improves the level of intelligent operation and maintenance management. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, a dynamic airflow control method for negative pressure wards in green and smart hospitals based on BIM is provided.

[0008] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: In a first aspect, a dynamic airflow control method for negative pressure wards in a green and intelligent hospital based on BIM is characterized by the following steps: Step S1: Construct the basic BIM model, including: creating a BIM model of the target negative pressure ward area, exporting the BIM model to IFC format and importing it into the digital twin platform, associating IoT sensor data with BIM model components through data mapping rules to form the initial state of the digital twin of the negative pressure ward. Step S2: Deploy an IoT sensing network and achieve real-time data fusion, including: deploying a multi-source sensor network within the physical space of the negative pressure ward to collect dynamic operational data in real time; uploading the dynamic data to the digital twin platform via the MQTT protocol; and performing timestamp alignment, spatial coordinate matching, multi-source state estimation, anomaly identification, and data repair on the dynamic data; and synchronously mapping and driving the digital twin with the processed real-time data; wherein the dynamic data includes environmental parameters, pressure parameters, equipment operating status, and personnel and behavioral events; Step S3: Set dynamic control rules and multi-objective optimization functions, including: Based on the sensing control specifications and green building energy-saving targets, define dynamic control rules using the event-condition-action paradigm, with maintaining the target pressure gradient as the core constraint, and with minimizing the total energy consumption of the ventilation system, minimizing the air age in key indoor areas, and maximizing the efficiency of pollutant dilution and removal as optimization sub-objectives, and establish a quantifiable multi-objective optimization function; Step S4: CFD simulation-based control strategy pre-simulation and intelligent decision-making, including: when the system detects the occurrence of dynamic triggering events or the continuous deviation of key differential pressure parameters from the safety threshold, the fast computational fluid dynamics simulation engine is automatically invoked, using the real-time sensor data processed in step S2 as the initial and boundary conditions of the simulation, and several candidate control schemes are simulated and deduced in the digital twin model at the second level, and the predicted results of total energy consumption, air age, and pollutant dilution and removal efficiency of each scheme are output; Step S5: Execute the optimal control command and implement feedback correction, including: comparing and analyzing the simulation results of each pre-simulation scheme in step S4, selecting the scheme with the best comprehensive performance according to the multi-objective optimization function, generating a structured control command set, sending it to the building automation control system to drive the actuator through the BACnet protocol, monitoring the system response in real time through the sensor network in step S2 and sending the feedback data back to the digital twin, and automatically triggering a new round of simulation and control optimization if the feedback effect does not meet expectations. Step S6: 3D visualization monitoring, early warning and data analysis, including: real-time rendering and display of pressure gradient cloud map, dynamic airflow streamline, pollutant diffusion simulation animation and equipment operation status in the BIM-based 3D visualization monitoring interface; when an anomaly is detected or a risk of infection and control is predicted, an audible and visual alarm is activated and alarm information is pushed to the mobile terminal; the entire life cycle operation data is recorded synchronously; and energy efficiency analysis reports and event traceability reports are supported.

[0009] Furthermore, the data mapping rules in step S1 include: creating a digital twin proxy object with a unique identifier and three-dimensional spatial coordinates for each physical sensor in the BIM model, establishing a mapping table between physical device IDs and globally unique identifiers of BIM components, and realizing the spatial association between sensor data and model components.

[0010] Furthermore, the real-time data fusion in step S2 includes the following steps: (1) Timestamp alignment: All sensor data are timestamped with a unified time based on the network time protocol, and data with different sampling frequencies are aligned to a unified time reference by using a sliding time window and resampling technology; (2) Spatial coordinate matching: Through the pre-calibrated spatial coordinate transformation matrix, the physical installation coordinates of all sensors and the real-time positioning coordinates of personnel are uniformly transformed into the global coordinate system of the BIM model; (3) Multi-source state estimation: For key state variables, the Kalman filter algorithm or its variant is used to fuse the readings of multiple sensors of the same or different types, filter out measurement noise, and obtain the optimal estimate. (4) Anomaly identification and data repair: The rule engine combined with statistical methods is used to identify abnormal data. For instantaneous anomalies, the filtering algorithm is used to smooth them. For multiple short-term missing data, interpolation or prediction models are used to fill them. For long-term faults, the equipment alarm is triggered and the device is switched to the preset safety mode.

[0011] Furthermore, in step S2, the collection of personnel and behavioral events adopts a fusion positioning scheme that combines ultra-wideband and video recognition. Ultra-wideband provides absolute coordinates, while video recognition provides visual contextual information. The data are then fused using a Kalman filter algorithm to obtain the number of personnel, their real-time location, and their movement trajectory.

[0012] Furthermore, the core constraints in step S3 include the following quantitative constraints: The pressure difference between the ward and the external corridor is maintained within the range of -15Pa to -30Pa, and the directional pressure gradient of ward pressure < buffer room pressure < external corridor pressure is satisfied. The air exchange rate should meet the specifications for the corresponding negative pressure ward or contaminated area; preferably ≥12 times / hour. Noise level ≤ 40 dB(A); The airflow velocity at the air outlet is ≤0.2m / s; Temperature range: 22-26℃; relative humidity range: 40-60%. A replacement alarm is triggered when the pressure difference across the HEPA filter reaches twice the initial resistance.

[0013] Furthermore, the dynamic control rules in step S3 are defined in the form of an event-condition-action paradigm. Each rule includes an event identifier, a trigger event type, additional logical conditions, and an execution action. The trigger event types include ward doors being continuously open for more than a threshold, the number of people in the room exceeding a threshold, and the pressure difference in the buffer room decreasing for more than a threshold. The thresholds are configurable parameters that can be adjusted by maintenance personnel through the management interface.

[0014] Furthermore, the multi-objective optimization function in step S3 is constructed using a weighted sum method, and the objective function is expressed as: F(X) = w1 * E'(X) + w2 * τ' avg (X) + w3 * (1 - η'(X)); Where X is a decision vector consisting of the frequency of the supply and exhaust fans and the opening degree of the dampers; E'(X) represents the normalized total energy consumption of the ventilation system: ; = Σ (a i * f i 3 ); in, and These represent the estimated maximum and minimum total energy consumption among all feasible control schemes under the current constraints; a i f represents the coefficient related to the characteristics of the wind turbine. i Indicates the frequency of the fan; τ' avg (X) represents the normalized air age of the patient's respiratory zone: ; ; in, Indicates the local air age at the monitoring point; , These are the estimated maximum and minimum values ​​of the corresponding parameter across all feasible options; η'(X) represents the normalized pollutant dilution efficiency: ; ; in, This represents the steady-state dissipation rate of pollutants within the ward. This represents the steady-state concentration of pollutants at the exhaust vent under control scheme X. w1, w2, and w3 represent weighting coefficients, and w1 + w2 + w3 = 1.

[0015] Furthermore, the fast computational fluid dynamics simulation engine in step S4 employs a reduced-order model technique to achieve second-level simulation and deduction. Specifically, it includes: in the offline stage, high-fidelity CFD software is used to simulate preset working conditions and build a pre-calculated sample library; in the online stage, interpolation algorithms or machine learning models are used to quickly predict the simulation results of new schemes, with a single prediction time ranging from milliseconds to seconds.

[0016] Furthermore, the generation of candidate control schemes in step S4 adopts a hybrid strategy that combines preset empirical rules with online intelligent search: several candidate schemes are predefined for high-risk events, and particle swarm optimization algorithm or genetic algorithm is used to search for better combinations of control parameters in the constraint space; the upper limit of the number of schemes is configured according to computing resources and real-time requirements, and the number of schemes in a single simulation is controlled within 5 to 10.

[0017] Furthermore, the structured control instruction set in step S5 is encoded in JSON format, and each instruction includes at least a device identifier, instruction type, target value, timestamp, priority, and transaction identifier; the actuator includes a variable frequency fan and an electric air valve, wherein the variable frequency fan is controlled by a 4-20mA or 0-10V analog signal, and the electric air valve is controlled by an analog output or a digital output signal.

[0018] Furthermore, the judgment of the feedback effect in step S5 not meeting expectations adopts a three-level evaluation mechanism: the first level evaluates the success of command execution, the second level evaluates the compliance of pressure gradient, and the third level evaluates whether the overall performance has reached the predicted value. If the first level evaluation fails, a retry mechanism is triggered. If the second level evaluation fails, a new round of control optimization is triggered. If multiple optimizations fail consecutively, the alarm level is upgraded to a higher level and the system switches to the preset safety mode. The maximum number of iterations for a single triggered event is 3.

[0019] Furthermore, the three-dimensional visualization monitoring in step S6 is implemented based on the WebGL technology stack. The pressure gradient cloud map is generated by the Kriging interpolation algorithm and superimposed on the indoor surface of the room with a semi-transparent texture. The dynamic airflow streamlines are generated by fourth-order Runge-Kutta method and displayed as a tubular model with directional arrows. The pollutant diffusion simulation is implemented using a particle system.

[0020] Furthermore, the early warning in step S6 adopts a three-level classification mechanism: Level 1 warning corresponds to negative pressure failure or serious exceedance of pollutants, triggering a buzzer, red flashing, pop-up window locking, and SMS push; Level 2 warning corresponds to pressure gradient deviation or filter blockage, triggering yellow flashing and APP push; Level 3 warning corresponds to abnormal energy consumption or equipment operating time exceeding the limit, triggering a blue prompt badge and summary push; Level 1 warnings can only be lifted after manual confirmation.

[0021] Furthermore, the full lifecycle operation data in step S6 adopts a hierarchical storage strategy: time-series data is stored in a time-series database and compressed using differential encoding and run-length encoding; the data of the most recent 30 days is stored on a solid-state drive as hot data; data older than 30 days is migrated to a mechanical hard drive as cold data; and permanently retained data is stored in partitions by year.

[0022] Secondly, the present invention provides a BIM-based green smart hospital negative pressure ward dynamic airflow control system for implementing the method, characterized in that it includes: The BIM digital twin module is used to build and store the BIM model of the target negative pressure ward area, forming a digital twin of the negative pressure ward. The Internet of Things (IoT) sensing module includes a multi-source sensor network deployed in the physical space of the negative pressure ward, used to collect dynamic operating data in real time, and to perform timestamp alignment, spatial coordinate matching, multi-source state estimation, anomaly identification, and data repair on the data. The rule engine and optimization module are used to define dynamic control rules using the event-condition-action paradigm, and to establish a multi-objective optimization function with maintaining the target pressure gradient as the core constraint and minimizing the total energy consumption of the ventilation system, the shortest air age, and the highest pollutant dilution efficiency as optimization sub-objectives. The CFD simulation and pre-simulation module is used to simulate and extrapolate candidate control schemes in seconds using real-time sensor data as initial and boundary conditions when a dynamic triggering event or pressure difference abnormality is detected, and outputs the prediction results of each scheme. The instruction execution and feedback module is used to generate a set of structured control instructions, which are sent to the building automation system via the BACnet protocol to drive the actuators, and closed-loop correction is performed using sensor feedback data. The 3D visualization monitoring module is used to render and display pressure gradient cloud maps, dynamic airflow streamlines, pollutant diffusion simulation animations and equipment operating status in a BIM-based 3D interface in real time, and trigger early warning push when abnormalities occur. The data storage and analysis module is used to synchronously record operational data throughout the entire lifecycle and supports the generation of energy efficiency analysis reports and event tracing reports.

[0023] The beneficial effects of this invention are: (1) A leap from static control to dynamic adaptation has been achieved. By deeply coupling the static BIM model with real-time IoT data and personnel behavior, the airflow control system can actively respond to various dynamic disturbances in actual operation, which significantly improves the reliability and robustness of infection control measures.

[0024] (2) It realizes the transformation from experience-based regulation to predictive optimization decision-making. It innovatively introduces a CFD rapid simulation link based on digital twins into the control closed loop, so that the regulation decision-making changes from passive response to ex-ante simulation and optimization selection, which greatly improves the accuracy, foresight and efficiency of regulation.

[0025] (3) It takes into account the synergistic optimization of infection control safety and operational energy efficiency. By establishing a calculation model with safety as a hard constraint and energy saving and other indicators as optimization goals, it automatically finds and executes more economical operation strategies under the premise of absolutely ensuring the negative pressure isolation effect, which is in line with the development concept of green and smart hospitals.

[0026] (4) It greatly improves the intuitiveness and efficiency of operation and maintenance management. The BIM-based three-dimensional visualization monitoring interface presents abstract environmental parameters and complex airflow organization in intuitive graphics, lowers the professional threshold, and qualitatively improves the speed of risk identification, location and emergency response. Attached Figure Description

[0027] Figure 1 This is a flowchart of the overall system of the method of the present invention.

[0028] Figure 2 This is a schematic diagram of the architecture for deploying a BIM-digital twin model and an IoT sensing network for a negative pressure ward. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0030] A BIM-based method for dynamic airflow control in negative pressure wards of green smart hospitals includes the following steps: S1: Constructing an integrated BIM digital twin foundation model Create a high-precision BIM model of the target negative pressure ward area. This model accurately includes the geometric information, spatial topology (such as room adjacency and doorway connectivity) of the building envelope, doors and windows, pass-through windows, air supply outlets, return / exhaust air outlets, ventilation ducts, and key medical equipment, as well as equipment attribute parameters (such as air outlet design air volume, filter efficiency rating, and fan performance curve).

[0031] The BIM model is then integrated with the IoT management platform, meaning the BIM model is exported in IFC format and imported into the digital twin platform. IoT sensor data is uploaded to the platform via the MQTT protocol. The platform uses self-developed data mapping rules to associate the data with model components. Finally, optimization instructions are sent to the building automation system via the BACnet protocol for execution, thereby constructing the initial digital twin of the negative pressure ward, which serves as a mapping of the physical entity in virtual space.

[0032] The self-developed data mapping rules are implemented as follows: In the digital twin platform, a corresponding virtual digital twin object is created for each physical sensor (such as a differential pressure sensor or a temperature and humidity sensor). This object has a unique ID in the platform and is assigned three-dimensional spatial coordinates (X, Y, Z) consistent with the physical installation location of the sensor, based on the global coordinate system of the BIM model. Simultaneously, the platform maintains a mapping table that records the association between the device ID (such as serial number or MAC address) of each physical sensor and the globally unique identifier (GUID) of its corresponding component (such as a room, duct, or equipment) in the BIM model. When sensor data is uploaded, the platform queries this mapping table to bind the data stream in real time and drive the corresponding component in the BIM model, thereby achieving a precise spatiotemporal association between data and the model.

[0033] S2: Deploy IoT sensing networks and achieve real-time data fusion The aforementioned "real-time data fusion" goes beyond simple "data mounting" (i.e., merely binding data to BIM components) in its technical implementation, and includes the following specific technical means: (1) Basic level (spatial association): First, it is indeed necessary to "mount", that is, to use "self-developed data mapping rules" to create a unique "digital twin" proxy object with three-dimensional spatial coordinates for each physical sensor in the BIM model. This is the basis for realizing "spatial dimension association".

[0034] (2) Advanced Level (State Estimation and Data Augmentation): For critical, multi-source state variables (such as pressure and temperature in a room), the system typically employs state estimation algorithms to obtain more reliable values ​​to drive CFD simulations. For example, for room pressure, Kalman filtering or its variants can be used. This algorithm can fuse readings from multiple sensors of the same or different types and effectively filter out measurement noise, estimating a smoother optimal value that is closer to the true state than any single sensor reading. This is a more robust and scientific fusion method than simple weighted averaging. In addition, for missing data points, interpolation algorithms (such as linear interpolation and spline interpolation) are used to fill in the gaps to form a complete time series.

[0035] The spatiotemporal alignment method involved is crucial for heterogeneous and asynchronous data. In other words, rigorous preprocessing is essential to utilize asynchronous, multi-frequency sensor data as boundary conditions in CFD simulations. This is a core step in constructing an effective digital twin. The specific implementation path is as follows: (1) Timestamp Alignment: All IoT sensors and data acquisition modules must add high-precision, unified timestamps based on network time protocols or other time synchronization mechanisms when collecting data. The data aggregation platform maintains a master timeline. For data streams with different sampling frequencies (such as 1Hz differential pressure and 10Hz positioning), the system will use sliding time windows and resampling techniques. For example, high-frequency data (10Hz) can be aligned to the time of low-frequency data (1Hz) by downsampling (such as taking the mean within the window), or all data can be unified to a higher or more synchronous time step required for CFD simulation by interpolation.

[0036] (2) Spatial coordinate matching: For example, to obtain the "real-time location" of personnel, this location data (such as UWB coordinates) needs to be aligned with the global coordinate system of the BIM model. This requires a pre-calibrated spatial coordinate transformation matrix to uniformly transform the physical installation coordinates of all sensors and the real-time positioning coordinates of personnel to the coordinate system of the BIM model. This ensures that the dynamic boundary condition of "personnel location" can be correctly used by the CFD model in three-dimensional space.

[0037] The involved sensor data anomaly identification and quality control mechanism. This is a regulatory decision to prevent "garbage in, garbage out," and the specific measures are as follows: (1) Anomaly Detection: The system needs to integrate a data quality control module. This module usually combines two aspects: ① Rule Engine: Detects anomalies (such as pressure difference anomalies) based on physical common sense and preset thresholds. For example, it detects whether the value exceeds the reasonable range, whether it remains unchanged for a long time (dead value), and whether a step jump occurs. ② Algorithm Model: It uses statistical methods (such as the 3σ principle to detect outliers) or more complex machine learning models (such as isolated forests) to identify latent anomalies that are difficult to describe with fixed rules, such as slow drift or complex noise.

[0038] (2) Data Repair and Processing Strategies: After identifying anomalies, the system will not discard the data directly (which would result in data loss), but will instead adopt strategies based on the anomaly type: ① Transient anomalies / noise: Smoothing is achieved using filtering algorithms (such as median filtering and low-pass filtering), or by replacing the data with interpolated valid data from before and after. ② Short-term missing data or continuous anomalies: The system can switch to "predictive mode" and use an autoregressive model based on historical data or a simplified physical model embedded in the digital twin model itself to generate alternative data, ensuring the continuity of the control loop. ③ Long-term faults: The system should trigger equipment fault alarms (such as alarm functions), notify maintenance personnel to perform repairs, and switch the relevant control loops to the preset safety mode.

[0039] (3) Confidence transfer: The processed data can be accompanied by a “data quality label” or “confidence score”. That is, in the data transmission structure, a “data_quality_label field” or a “confidence_score field” is added to the corresponding data point and passed to the downstream CFD simulation and decision-making modules. Confidence can be introduced as a weight so that these modules can know the reliability of the data and weigh it in the calculation.

[0040] A multi-source sensor network is deployed within the physical space of the negative pressure ward to collect dynamic operational data in real time. This real-time data is then synchronously mapped and used to drive the digital twin via a data interface (a software service based on the MQTT protocol for data access, including data parsing, format conversion, preliminary cleaning, and internal writing functions). The dynamic data includes at least: (1) Environmental parameters: temperature, relative humidity, carbon dioxide concentration, particulate matter concentration, and volatile organic compound concentration in each functional area (such as wards, buffer rooms, and toilets).

[0041] (2) Pressure parameters: real-time pressure difference between each room and the external corridor or adjacent clean area, and pressure difference on both sides of the key door gap.

[0042] (3) Equipment operating status: operating frequency, start / stop status, and fault signals of the blower and exhaust fan; opening degree of the air valve; pressure difference across the high-efficiency filter.

[0043] (4) Personnel and Behavioral Events: The number of people, their real-time location, and movement trajectories indoors are obtained through UWB, RFID, or video recognition technologies; the door opening and closing status and duration are monitored through door magnetic sensors. A high-reliability, high-precision solution is preferred: UWB and video recognition are deployed in parallel to complement each other, and fused using Kalman filtering algorithms to obtain stable, accurate, and reliable personnel location information to drive subsequent intelligent control. UWB provides absolute coordinates but may be affected by multipath interference; video recognition provides rich visual context and identity information but is greatly affected by lighting and occlusion; RFID can serve as an auxiliary to UWB for precise triggering and identity binding at key access points.

[0044] S3: Define dynamic control rules and multi-objective optimization functions Based on national infection control standards such as the "Environmental Control Requirements for Negative Pressure Isolation Wards in Hospitals" and green building energy conservation goals, core control logic is pre-defined in the BIM-digital twin platform: Core constraints: The pressure in the contaminated area (ward) must be kept at the lowest level, and the pressure gradient must be kept stable (i.e., ward pressure < buffer room pressure < external corridor pressure). The specific quantitative constraints are as follows: ① Pressure gradient target: The pressure difference between the ward and the external corridor should be maintained within the range of -15Pa to -30Pa as the "core safety constraint"; ② Air exchange rate: It must meet the mandatory requirements of the "Environmental Control Requirements for Negative Pressure Isolation Wards in Hospitals" and other standards for the minimum air exchange rate in contaminated and semi-contaminated areas (such as the usual requirement of ≥12 times / h), which is the basis for ensuring the dilution and removal of pollutants; ③ Noise level: It must meet the background noise limit requirements of the "Code for Sound Insulation Design of Civil Buildings" and "Technical Code for Clean Operating Rooms in Hospitals" for areas such as wards and buffer rooms (such as ≤40dB(A) at night in wards) to ensure patient rest and a quiet medical environment; ④ Temperature and humidity range: It must maintain the indoor temperature and humidity range required for patient comfort and medical operations (such as temperature 22-26℃, relative humidity 40-60%); ⑤ Wind speed and draft sensation: The airflow speed at the air outlets must be limited to avoid creating an unpleasant draft sensation in the patient activity area (controlled at 0.2). ⑥ Filter resistance and lifespan: The system needs to monitor the pressure difference across the HEPA filter. The initial resistance is usually between 100 Pa and 250 Pa, and the replacement pressure difference (final resistance) is usually set to twice its initial resistance. For example, if the initial resistance is 150 Pa, the replacement alarm threshold can be set to 300 Pa.

[0045] Dynamic event triggering mechanism: Specific personnel behaviors and equipment status changes collected in step S2 are defined as control trigger events. The implementation paradigm preferentially adopts the "Event-Condition-Action (ECA) rule". For example (setting parameters): Event: door_open_duration_exceeds (the ward door remains open for longer than the threshold); Condition: door_id == "Ward_001_MainDoor" && duration > threshold (the door ID is "Ward 001 Main Door" and the duration is greater than the threshold); Action: trigger_CFD_simulation() (triggers CFD simulation), etc. Specific threshold examples for some events are as follows: "ward door open for more than 5 seconds", "real-time number of people in the ward ≥ 3", "pressure difference in the buffer room higher than -5Pa", etc. These thresholds (such as 5 seconds, 3 people, -5Pa) are input into the rule engine as configurable parameters, not fixed values. The platform provides a management interface that allows maintenance personnel to adjust and set these thresholds based on factors such as the specific purpose of the ward (e.g., ordinary negative pressure ward vs. intensive care negative pressure ward), actual size, and risk level, in order to meet the personalized management needs of different wards.

[0046] Multi-objective optimization function: With maintaining the target pressure gradient as the core constraint, and minimizing the total energy consumption of the ventilation system, minimizing the air age in key indoor areas (such as the patient's breathing zone), and maximizing the efficiency of pollutant dilution and removal as optimization sub-objectives, a quantifiable multi-objective optimization model is established.

[0047] The aforementioned multi-objective optimization function constructs a quantitative optimization model based on the weighted sum method. Triggered by events or state anomalies, it utilizes a CFD reduction model to guarantee a solution within seconds, and forms a hierarchical collaboration with dynamic rules through "rule triggering and optimization decision-making." Further explanation includes the following three aspects: (1) Regulations on mathematical expressions, quantification, normalization and weighting.

[0048] To transform the concepts in the documentation into computable and solvable engineering problems, a typical multi-objective optimization model can be constructed as follows: ① Decision variables: These are usually adjustable control parameters in the system, such as the frequency f_i of each supply and exhaust fan, the opening degree v_j of each air valve, etc., which constitute the decision vector X = [f_1, f_2, ..., v_1, v_2, ...].

[0049] ② Core constraints: hard constraints already defined in Document S3, such as: pressure gradient: -30 Pa<= P_ward - P_corridor<= -15 Pa and P_ward<P_buffer<P_corridor. Other engineering constraints: air change rate ACH>=12, noise L_A<= 40 dB(A), wind speed v_local<= 0.2 m / s, etc.

[0050] ③ Quantification of sub-objective functions: a) Minimum total energy consumption: Energy consumption E can be modeled as a function of fan frequency, and a commonly used approximation formula is E = Σ (a_i * f_i^3), where a_i is a coefficient related to fan characteristics. The objective is Min E(X). b) Shortest average air age: Air age τ is an indicator for measuring air freshness. A number of key points can be set in the patient's breathing area, and its average air age τ_avg can be estimated through CFD simulation or a simplified model (such as turbulent ventilation model). The objective is Min τ_avg(X). c) Highest pollutant removal efficiency: It is usually measured by average pollutant concentration C_avg or removal efficiency η, which can be calculated from the results of pollutant transport equation obtained by CFD simulation. The objective can be Min C_avg(X) or Max η(X).

[0051] ④ Multi-objective processing: Since sub-objectives usually conflict with each other (e.g., reducing energy consumption may increase air age), multi-objective optimization methods are required. A common practice is to construct a weighted sum single objective function or adopt Pareto optimization. a) Weighted sum method: Convert multiple objectives into a single objective: F(X) = w1 * E'(X) + w2 * τ'_avg(X) + w3 * (1 - η'(X)). Where, E'(X), τ'_avg(X), η'(X) are sub-objective functions after normalization (for example, divide by their design reference value or extreme value within the feasible range, so that they are dimensionless and have similar orders of magnitude). The weight coefficients w1, w2, w3 (w1 + w2 +w3 = 1) reflect the relative importance attached to energy consumption, air quality and infection control efficiency, and need to be preset by domain experts or according to the hospital's operation strategy. b) Weight determination: Weight coefficient is a key engineering decision parameter, and its determination method is not specified in the document. In practice, it can be determined through analytic hierarchy process, expert scoring method, or combined with historical operation data and risk assessment.

[0052] The normalized total energy consumption E′(X) of the ventilation system is obtained by normalizing the original total energy consumption E(X). A typical normalization method is: , wherein, and These represent the estimated maximum and minimum total energy consumption among all feasible control schemes under the current constraints. These extreme values ​​can be predetermined using historical operating data, design conditions, or an offline CFD simulation sample library. Through this process, It was scaled to be near the [0, 1] interval. The smaller the value, the better the energy efficiency.

[0053] The average air age of the patient's respiratory zone The calculation is performed through the following steps: First, N representative monitoring points are defined within the patient's breathing zone (typically defined as a spatial area within a height range of 0.6-1.2 meters above the bed). Then, using the aforementioned fast CFD simulation engine, the local air age at these monitoring points is calculated under control scheme X. Finally, the arithmetic mean of these local values ​​is taken as the representative value for the region: Among them, local air age The air age can be obtained in a CFD simulation by solving an additional air age transport equation, or estimated using a simplified model (such as a turbulent ventilation model). Air age is a scalar quantity characterizing the time air has resided in a room. In a CFD simulation, it is obtained by solving an additional scalar transport equation. That is, the air age transport equation: Where: τ is the local air age (s), ρ is the air density (kg / m³), u is the velocity vector (m / s), and is the effective diffusion coefficient of the air age (kg / (m·s)), usually determined by a turbulence model (such as k... -ε The model determines that it is related to the turbulent diffusivity, i.e. Where μ is the effective dynamic viscosity, It is the turbulent Schmidt number of the air age (usually taken as around 1.0). Convection term, ρ is the diffusion term, and ρ is the source term. Its physical meaning is that the "age" of the air increases naturally at a rate of 1 second per second. This is a unique, always positive source term in the air age equation. Boundary conditions for the air age transport equation: the air age at the air outlet is 0 seconds (fresh air), and zero gradient boundary conditions are typically used for the walls and solid surfaces. Exhaust vents typically employ outflow boundary conditions. Solving this equation in conjunction with the continuity equation, momentum equation, energy equation, and turbulence equations (such as the k-ε equation) yields the air age distribution τ(x,y,z) at various points in space. The average air age τavg in the patient's breathing zone is obtained by spatially averaging the τ values ​​within that region (e.g., defining N monitoring points within a space 0.6-1.2 meters above the bed), i.e.: Traditional high-fidelity CFD solutions to the entire set of equations (including the air age equation) take several hours, failing to meet real-time optimization requirements. The reduced-order model (ROM) employed in this invention achieves a sub-second response time through an "offline training, online prediction" model. Its core lies in moving the time-consuming CFD physics solution process offline and condensing the resulting knowledge into a fast input-output prediction model (ROM). During online optimization, the main processes include real-time status input, rapid ROM query and calculation, support for multi-objective optimization, and optimization iteration. Only lightweight calculations or queries are needed on this ROM, thus bypassing the massive computational load of traditional CFD simulations and achieving the sub-second response time required for real-time multi-objective optimization. This method ensures that the key sensory and comfort indicators such as air age and pollutant concentration used for optimization are still based on high-fidelity physical principles, rather than empirical formulas, achieving the optimal engineering balance between speed and accuracy.

[0054] The pollutant dilution efficiency This is used to quantify the ability of a ventilation system to remove pollutants. One definition is based on the pollutant removal efficiency under steady-state conditions: , in, The steady-state emission rate of pollutants in the ward (or set as a reference concentration). The steady-state concentration of pollutants at the exhaust vent under control scheme X. The pollutant transport equations can be obtained through CFD simulation. The closer the value is to 1, the better the pollutant removal effect. The pollutant transport equation, which satisfies the law of conservation of mass in the transport of pollutants in the ward air, is typically treated as a passive scalar (its concentration does not affect the main airflow) in CFD simulations. Its spatial distribution and temporal evolution are obtained by solving an additional scalar (i.e., pollutant concentration) transport equation. The general pollutant transport equation is: Where Y is the mass fraction of the pollutant (dimensionless), or can be replaced by C to represent the concentration (e.g., kg / m³). Under steady-state and constant density conditions, the equation is often simplified to a form with the concentration C as the variable, where ρ is the air density (kg / m³), u is the velocity vector (m / s), and D is the effective mass diffusion coefficient of the pollutant (m² / s). This is the convective term, indicating that pollutants move with the airflow. The diffusion term represents the diffusion of pollutants due to their concentration gradient, while S is the source term (e.g., kg / (m³·s)) representing the rate of pollutant generation or disappearance. For simulating pathogen aerosols exhaled by patients, S is typically set to a constant release rate S0 in the patient's breathing zone (e.g., near the mouth and nose). This invention differs from traditional high-fidelity CFD simulations (which take several hours to solve the above equations). The reduced-order model establishes the input (control scheme) and output (e.g., C) in an offline phase. out A fast mapping relationship between C and τavg is established to achieve second-level prediction. Its implementation path is consistent with the principle of obtaining the air age τavg, both following the "offline training, online prediction" model. Within the framework of this invention, "C is obtained within seconds using a reduced-order model." out The technical essence of this is to transform the high-fidelity CFD physical solution process, which includes pollutant transport equations, into a lightweight input-output surrogate model trained on a large number of offline samples.

[0055] The normalized air age of the patient's respiratory zone and normalized pollutant dilution efficiency A normalization method similar to that used for energy consumption is employed for processing: , ,in, , , , These are the estimated maximum and minimum values ​​of the corresponding parameters across all feasible solutions, respectively. Through normalization, The smaller the value, the fresher the air. The larger the value, the higher the pollutant removal efficiency. In the weighted sum function F(X), the efficiency term is treated with (1- The form of ) is to be consistent with energy consumption and air age, meaning that the smaller the value, the better the performance.

[0056] (2) Regarding the optimization of triggering timing, computational complexity and real-time performance.

[0057] ① Triggering Timing: Optimization is primarily performed in two scenarios: a) Event Trigger: It is triggered immediately when a dynamic event defined in S3 (such as door opening timeout or personnel exceeding limits) occurs. b) Periodic / Status Trigger: The system self-checks core parameters (such as differential pressure) at regular intervals (e.g., every minute). If it detects a continuous deviation from the safety threshold, optimization is triggered even without external events to perform preventative adjustments. This avoids the lag of responding only after an incident occurs.

[0058] ② Sources of computational complexity: The complexity of the optimization problem mainly stems from CFD simulation. Performing high-precision CFD calculations on complex three-dimensional spaces like negative pressure isolation wards using traditional methods can take several hours, failing to meet real-time control requirements.

[0059] ③ Real-time performance guarantee: First, a pre-computation sample library is built. Before system deployment, high-fidelity CFD offline calculations are performed on a large number of typical operating conditions (different supply and exhaust air combinations, door states, personnel positions), and the results (such as pressure difference, air age, and concentration field) are stored in a database. During online optimization, the effect of the new control scheme is predicted within milliseconds using interpolation or machine learning-based regression models. Then, the physical model is simplified using simplified models based on flow network methods or region methods, sacrificing some spatial details for a significant increase in computational speed. The aforementioned fast simulation engine, with its optimization algorithms (such as gradient descent, genetic algorithms, and particle swarm optimization), can typically complete the search within seconds when the dimensionality of decision variables is low (a few to dozens of fans and valves), thus meeting the requirements of closed-loop real-time control.

[0060] (3) On the synergistic relationship between dynamic control rules and multi-objective optimization functions.

[0061] Within the system framework described in the document, there is a clear, hierarchical collaborative logic of rule priority and optimization assistance between dynamic control rules and multi-objective optimization functions: ① The primacy and safety of rules: The rules defined in the "Dynamic Event Triggering Mechanism" (especially ECA rules) are the first line of defense and direct triggers for the system response. Its design goal is to respond quickly and deterministically to known, high-risk events to ensure the bottom line of infection control safety. For example, once the event "ward door open for more than 5 seconds" is triggered, the system will immediately execute preset, validated emergency actions (such as triggering simulation and control in S4). This response is mandatory and has the highest priority. Rules ensure the system's immediate response capability in emergency situations.

[0062] ② Optimization Precision and Economy: The "multi-objective optimization function" is a refined tool used to decide how to implement control measures after a rule is triggered. When a rule triggers a CFD simulation (as described in S4), the optimization model calculates and selects the optimal overall performance scheme from among numerous feasible control options (schemes A, B, and C). It weighs multiple objectives such as energy consumption and air quality, striving to achieve more economical and comfortable operation while meeting safety constraints (core constraints).

[0063] ③ Explanation of Collaborative Logic: a) Relationship: Rules are the "switch" and "trigger," while optimization is the "calculator" and "selector." Rules determine "when and why" to initiate the optimization decision-making process, while optimization determines "how and which" control action to execute. b) Process: Dynamic event / state anomaly → trigger rule → call fast CFD simulation engine → evaluate candidate solutions based on multi-objective optimization function → select optimal solution → execute control. This is a sequential, rule-first process. c) Coverage: In the document description, optimization results do not directly rewrite pre-defined rules. Rules (including triggering conditions and thresholds) are pre-configured, relatively stable business logic. Optimization, within the framework of rule triggering, performs real-time calculations for each specific control action. Both perform their respective functions, together forming a complete intelligent decision-making system.

[0064] S4: CFD Simulation-Based Control Strategy Prediction and Intelligent Decision Making When the system detects a dynamic triggering event or finds that key differential pressure parameters continuously deviate from the safety threshold through periodic self-checks, it automatically invokes the fast computational fluid dynamics simulation engine integrated with the platform. Its features are described below: (1) Regarding the types of simulation engines, integration methods, and data flow paths.

[0065] The specific technical architecture involves an engine core based on a self-developed fast solver (such as using the region method or pre-computation models), or a deeply customized and encapsulated open-source solver (such as OpenFOAM). This system pre-generates a large number of offline calculations for specific ward geometries to generate a sample library, and develops corresponding online interpolation and prediction interfaces. Integration with the platform is typically achieved through application programming interfaces (APIs) or microservices.

[0066] The data flow path can be summarized into four steps: ① Triggering and initialization: After an event is triggered or a periodic self-check is performed, the platform (rule engine) calls the simulation engine API; ② Input data transmission: The platform transmits the real-time data processed by S2 (environmental parameters, pressure parameters, equipment status, personnel location and behavior after fusion, alignment and quality control) as the initial and boundary conditions for the simulation, along with the current geometric and topological state of the BIM digital twin, to the simulation engine; ③ Scheme simulation: The simulation engine performs rapid simulation and deduction on several candidate control schemes (schemes A / B / C, i.e., different combinations of fan frequencies and valve openings); ④ Result return: The simulation engine returns the prediction results of each scheme (three-dimensional spatial pressure distribution, airflow organization, pollutant concentration trends, etc.) to the platform's core decision-making module.

[0067] (2) Regarding the stopping conditions, convergence criteria and accuracy guarantee of the simulation.

[0068] The convergence control and accuracy guarantee mechanisms of the aforementioned "fast" simulation engine differ from those of traditional high-fidelity CFD, primarily relying on its underlying reduced-order model technology: ① Convergence of the reduced-order model: If the engine uses a pre-computed sample library + interpolation / machine learning model, its "simulation" is essentially querying and computation, without an iterative convergence process. Its "accuracy" is predetermined when the sample library is generated offline. The speed in the online phase is instantaneous (milliseconds). ② Convergence of the simplified model: If a simplified model such as the flow network method is used, its solution process may involve iterating linear equations, but the scale is much smaller than that of CFD. Its stopping condition is usually that the residual is below a certain preset threshold (e.g., 1e-4). Due to the low dimension of the equations, it can also be completed within milliseconds to seconds.

[0069] The accuracy guarantee mechanism is primarily established during the offline phase before system deployment. Key components include: ① Model validation: When generating the sample library or calibrating simplified model parameters, high-fidelity CFD software (such as ANSYS Fluent) must be used to perform detailed simulations under numerous typical operating conditions. The results are then used as the "gold standard" to train and validate the accuracy of the reduced-order model or correct the simplified model. For example, ensuring that the error between the fast engine's results and the high-fidelity simulation results is less than an acceptable range (e.g., 5%) for key prediction indicators (such as ward pressure differential and respiratory zone wind speed); ② Scope definition: Clearly define the applicable operating conditions range for the fast engine (e.g., 1-4 people, door opening 0-90 degrees) and set boundaries in the platform logic. When real-time operating conditions exceed this range, trigger special handling (e.g., alarm and switch to safe mode, or start a complete high-fidelity simulation); ③ Online verification (optional): Periodically (e.g., once a day) when the system is idle, use real-time data as boundaries to simultaneously run the fast engine and the high-fidelity engine for comparison, continuously monitoring the drift of the fast model.

[0070] (3) Regarding the simulation execution and resource allocation when multiple events occur concurrently.

[0071] In real-world scenarios, consider ensuring "second-level" response times under high concurrency. A complete system design should consider the following aspects: ① Event Priority and Queuing: Not all events need or should trigger time-consuming simulation optimizations immediately. The rules defined in S3 should include event priorities. For example, the priority of a device fault alarm such as "high-efficiency filter differential pressure exceeding the limit" may be higher than "temporary increase in personnel." The system can use a priority queue to manage triggered tasks, sorting them by trigger time within the same priority. ② Simulation Task Scheduling: Optimization uses parallel processing. If the platform's computing resources (such as the number of server cores) are sufficient and the simulation engine supports it, simulation tasks for different physical ward units can be processed in parallel, as they are usually independent of each other. Serial processing can also be performed. For multiple events triggered in a short period within the same ward unit (such as someone entering while the door is not closed), the system can adopt a "latest state overwrite" strategy. That is, if the second event is triggered before the simulation of the first event is completed, the boundary conditions of the ongoing simulation task are updated with the latest real-time state, or the previous unfinished task is canceled, and a new simulation is started directly based on the latest state. This avoids making decisions based on outdated states and saves resources. ③ Computational Resource Guarantee: To achieve stable "second-level" response, resource estimation and reservation are first required. Based on the number of wards, the maximum expected frequency of events, and the duration of a single simulation, the required computing resources (CPU, memory) are estimated and reserved or elastically scaled in the server cluster. Secondly, containerization and microservices are implemented, encapsulating the simulation engine as a Docker container for rapid deployment, scaling, and resource isolation. Multiple simulation engine instances are managed using orchestration tools such as Kubernetes to handle high concurrency. Finally, a timeout and degradation mechanism is implemented, setting a timeout (e.g., 2 seconds) for each simulation task. If no result is returned after the timeout, optimization is abandoned, and a pre-defined, verified, and conservative contingency plan is executed to ensure that the real-time performance and security of the system control loop are not compromised by computational delays.

[0072] Using the real-time sensor data obtained in step S2 as the initial and boundary conditions for the simulation, several candidate control schemes are simulated and extrapolated in the digital twin model at a second-level. These candidate schemes include adjusting the airflow at specific supply / exhaust vents, adjusting the frequency of the variable frequency fan, changing the opening of relevant dampers, and adjusting the pressure setpoints of adjacent areas in a coordinated manner. The simulation outputs the predicted results after each scheme is executed, including the three-dimensional spatial pressure distribution, airflow organization morphology, and pollutant concentration diffusion trend.

[0073] The "initial and boundary conditions of the simulation" are set as follows: the initial conditions are the environmental conditions inside the space (temperature, humidity, pollutant concentration, and personnel distribution); the boundary conditions are the inputs that drive changes in the system, mainly the equipment operating parameters (airflow at vents, fan status) and the physical opening status (door opening and closing). All data are derived from real-time, high-quality sensor data processed by S2.

[0074] The key to the aforementioned "second-level simulation" is not hardware acceleration of traditional CFD solvers, but rather the fundamental adoption of reduced-order model techniques. Specifically, it simplifies the physical model through pre-computed sample libraries and interpolation / machine learning predictions, or methods such as flow network methods / region methods, avoiding online high-fidelity CFD solutions, thereby achieving a performance leap from "hour-level" to "second-level".

[0075] The aforementioned pre-computation sample library combined with interpolation / machine learning prediction is the core technology for achieving "second-level" response. In the offline phase, the system uses high-fidelity CFD software (such as ANSYS Fluent) to simulate a massive number of preset working conditions, forming a vast "input (boundary condition combination) - output (flow field result)" database. In online applications, when a new scheme needs to be evaluated, the simulation engine does not perform a true CFD solution. Instead, it uses this new scheme as "input," searches for the nearest neighbor sample points in the pre-computation library, and directly and quickly predicts the "output" result using interpolation algorithms or machine learning models trained on these samples (such as neural networks or Gaussian process regression). This process is essentially data querying and computation, thus achieving "millisecond-level" response.

[0076] The aforementioned simplified physical models, such as the flow network method and the region method, simplify the room into nodes and doors, windows, and air vents into resistance elements connecting these nodes, using a set of algebraic equations to describe the pressure-flow relationship. The region method divides the room into several macroscopic, uniform regions. Both methods completely avoid solving the three-dimensional Navier-Stokes equations, transforming the complex CFD problem into solving a low-dimensional system of linear or nonlinear equations. The computation speed is several orders of magnitude faster than traditional CFD, while still meeting the requirement of "second-level" response.

[0077] The generation of candidate solutions employs a hybrid strategy combining "pre-defined empirical rules" and "online intelligent search." Pre-defined contingency plans ensure a rapid and reliable response to high-risk events; the online optimization algorithm offers the potential to find more refined and economical solutions while meeting safety constraints. No upper limit is specified for the number of solutions; this is a configurable parameter in practical systems, depending on computational resources and real-time requirements, and is typically limited by the number of iterations of the optimization algorithm or the population size. It is preferable to keep the number of solutions within 5 to 10.

[0078] The simulation output of the predicted results after the execution of each scheme involves a quantitative evaluation index system for the simulation results. The evaluation index is directly related to the multi-objective optimization function: the core of document S3, the "multi-objective optimization function," and its construction description constitute a complete quantitative evaluation index system. It explicitly sets "minimum total energy consumption of the ventilation system," "shortest air age in key indoor areas," and "highest efficiency in pollutant dilution and removal" as sub-objectives that need to be quantified and optimized. Specific quantification methods are detailed in the "Mathematical Expressions, Quantification, Normalization, and Weights" section of S3. For schemes that pass the constraint test, the system uses the "multi-objective optimization function" defined in S3, F(X) = w1*E'(X) + w2*τ'_avg(X) + w3*(1-η'(X)), for calculation. Here, E'(X), τ'_avg(X), and η'(X) are normalized sub-objective values, and w1, w2, and w3 are pre-set weights. The calculated F(X) value is the quantitative comprehensive score for each scheme. The system automatically selects the solution with the smallest comprehensive score F(X) (i.e. the lowest comprehensive cost) as the optimal solution.

[0079] S5: Execute the optimal control command and achieve feedback correction. By comparing and analyzing the simulation results of each pre-simulation scheme in step S4, the scheme with the best overall performance (e.g., minimum energy consumption increase and shortest pressure recovery time) under the premise of satisfying the core pressure gradient constraint is selected. The platform automatically generates a structured control command set and sends it to the building automation system through a standard interface to drive the corresponding variable frequency fans, electric dampers, and other actuators. After the control action is executed, the system response is monitored in real time through the sensor network in step S2, and the feedback data is transmitted back to the digital twin, forming a closed loop of "perception-decision-execution-evaluation". If the feedback effect does not meet expectations, the system automatically triggers a new round of simulation and control optimization.

[0080] In modern building automation and IoT systems, the structured control instruction set typically employs a lightweight, highly readable, and easily transmitted standardized data format. The two most common are JSON and XML, with JSON being more prevalent in IoT and Web services due to its simplicity. A complete control instruction usually includes at least the following key fields to ensure parsability by downstream systems: ①device_id (device ID): A unique device identifier defined in the BIM model, used to accurately locate and control targets in the building control system.

[0081] ②command_type: such as SET_SPEED (set frequency), SET_POSITION (set opening degree), SET_STATE (set start and stop).

[0082] ③target_value: The specific value of the command, such as fan frequency (unit: Hz), damper opening degree (unit: % or angle), equipment status (0=closed, 1=open).

[0083] ④timestamp: The time when the command was generated, used for logging and event sorting.

[0084] ⑤priority: Indicates the urgency of the instruction and can be used for scheduling in scenarios with a sudden surge of instructions.

[0085] ⑥validity_duration: Optional, indicates how long the instruction is valid.

[0086] ⑦transaction_id (transaction ID): Uniquely identifies a control transaction, facilitating association with feedback data and used for the "feedback effect" evaluation described in S5.

[0087] The standard interface, as explicitly described in step S1, uses a communication protocol for issuing commands: "The optimized command is then sent to the building automation system for execution via the BACnet protocol." BACnet is the most widely used international standard communication protocol in the field of building automation. Based on this, its interface calling method, authentication, and timeout mechanism are typically as follows: ① Invocation method: The platform acts as a BACnet client, and the field controllers (such as DDCs) of the building automation system act as BACnet servers. The platform modifies the "current value" property of the controlled object (such as the fan frequency object) by calling the BACnet standard read and write property services (such as WriteProperty), thereby achieving control.

[0088] ② Authentication Mechanism: The BACnet protocol itself supports optional device object passwords and message source authentication. In actual engineering, it is more common to implement security controls at the network layer (such as VLAN isolation, firewall policies) and application layer (BACnet device instance number, IP address filtering).

[0089] ③ Timeout and Retry: The platform sets an operation timeout (e.g., 2-5 seconds) when calling the BACnet service. If no confirmation response is received within the timeout period, the platform can retry a limited number of times (e.g., 1-2 times) according to a preset strategy. If the retry still fails, the alarm described in S6 will be triggered, and the system may switch to a backup control strategy.

[0090] The aforementioned actuator action represents a typical, standardized process completed by a building automation system, converting digital commands into physical signals. The typical path for converting commands into actions is as follows: ① Command reception: The building automation system (via the BACnet interface) receives the command set from the platform.

[0091] ② Command parsing and routing: The system's central or field controller parses commands and finds the corresponding control logic based on the device_id.

[0092] ③ Signal Output: For variable frequency fans: The most typical control method is to generate a standard 4-20mA or 0-10V signal through the analog output (AO) module inside the controller, which is then sent directly to the analog input terminal of the frequency converter to precisely set the fan speed. Some newer fans may also receive the speed setpoint directly through communication protocols such as Modbus RTU / TCP. For electric dampers: Typically, the controller controls the valve actuator's positioning through an analog output (AO) signal (4-20mA / 0-10V) to achieve continuous adjustment from 0-100%. For two-way valves that only require switching, control is achieved through digital output (DO).

[0093] ④ Response Time: The total time from the issuance of a command to the actuator's action depends on the controller's scan cycle, network latency, and the actuator's mechanical inertia. For process control such as pressure regulation, an overall response time of several seconds to less than a minute is generally acceptable. The embodiment in document S5 mentions contingency plan B, "restoring the pressure gradient to a safe range within 3 seconds," which means the total time from the issuance of a command to the system pressure reaching a new steady state is on the order of seconds. This requires the actuators (fans, valves) themselves to have a relatively fast adjustment speed.

[0094] If the feedback effect does not meet expectations, a new round of optimization will be triggered. To ensure this feedback loop runs automatically, explicit effect evaluation rules must be preset in the S3 rule engine or optimization model. These typically include: ① Evaluation indicators: These are the key parameters that need to be verified. The most crucial one is obviously the pressure gradient (such as the pressure difference between the ward and the corridor), and may also include the air volume of key air outlets, etc.

[0095] ② Target Value and Allowable Deviation: Set the expected target value (i.e., the result predicted by the optimization plan) of the above key parameters after the control action is executed, and an acceptable deviation range. For example, the expectation is to restore the ward pressure differential to -20 Pa, with an allowable deviation of ±2 Pa.

[0096] ③ Stabilization Waiting Time: After the control action is executed, the system needs a period of time to reach a new stable state. Therefore, a reasonable stabilization waiting time (or "observation period") needs to be set before data collection for effect evaluation begins. This time should be determined based on the system's inertia; in the second-level response scenario of the document's example, it may be 10-30 seconds.

[0097] ④ Judgment Logic: After the stabilization waiting period ends, the system collects sensor feedback data. If the actual value of the key parameter continuously (e.g., for 3 consecutive sampling cycles) exceeds the allowable deviation range of the expected target value, it is judged as "the effect did not meet expectations".

[0098] ⑤ Retry and Upgrade: If the expected results are not met, the system can automatically trigger a new round of optimization (as described in S5). Simultaneously, a maximum number of retries should be set. If multiple consecutive optimizations fail, the alert should be upgraded to a higher level (as described in S6), indicating potential equipment failure, model inaccuracy, or significant unforeseen interference, requiring manual intervention.

[0099] S6: 3D Visualization Monitoring, Early Warning and Data Analysis The BIM-based 3D visualization monitoring interface displays real-time pressure gradient cloud maps, dynamic airflow streamlines, pollutant diffusion simulation animations, and the operating status of all equipment. When the system detects abnormal pressure differences, filter blockage alarms, or predicts infection control risks through simulation, the platform immediately activates audible and visual alarms and pushes alarm information and handling suggestions to the mobile terminals of operation and maintenance management personnel. The platform synchronously records full lifecycle operation data, supports the generation of energy efficiency analysis reports and event tracing reports, and provides data support for continuous optimization of operation and maintenance strategies. Example

[0100] Taking a standardized negative pressure ward unit in a newly built infectious disease building of a hospital as an example, the method of this invention is applied.

[0101] S1 Implementation: A detailed BIM model including wards, buffer rooms, and individual restrooms was created using Autodesk Revit software. Equipment-level modeling was performed for components such as air supply ceilings, downdraft vents, airtight automatic doors, pass-through windows, and exhaust fan housings, adding attribute parameters such as equipment ID, design air volume, affiliated ventilation system, and filter specifications. This model was then imported into the "Smart Hospital Operation and Maintenance Digital Twin Platform" as the core digital foundation.

[0102] S2 Implementation: High-precision differential pressure sensors and temperature and humidity sensors are installed in the ceilings of wards and buffer rooms; ultrasonic air volume meters are installed in the main supply and exhaust ducts; door magnetic sensors are installed on the frames of all airtight doors; and medical staff are equipped with active RFID positioning work cards. All sensor data is collected through an industrial IoT gateway and transmitted to the platform. The platform, based on a preset mapping relationship, associates and integrates the data stream with the corresponding entity objects in the BIM model in real time in both spatial and temporal dimensions.

[0103] In the scenario described above, the core purpose of data fusion is to provide high-quality, reliable estimates of key state variables for CFD simulation and intelligent decision-making, rather than simply aggregating data. The following example uses the pressure difference between the ward and the external corridor—a key infection control parameter—to illustrate how to fuse multi-source data using the Kalman filter algorithm. The specific steps are as follows: Step 1: Establish the system state-space model First, a simplified dynamic model needs to be established for the state variable to be estimated (ward pressure differential). Since the pressure differential change is mainly caused by events such as fan operation and door opening and closing, it can be approximated as a random walk model with process noise over a short time scale. Meanwhile, the system has multiple observations (sensor readings). The dynamic model satisfies the following two equations: ① Equation of state: ,in, This represents the actual pressure difference at time k (to be estimated). It is process noise, assumed to follow a normal distribution with a mean of 0 and a covariance of Q, i.e. The Q-value reflects the uncertainty of the model's prediction of state changes.

[0104] ② Observation equation: ,in, This is the observation vector at time k. In this embodiment, it is assumed that two micro-differential pressure sensors are deployed (e.g., one inside the ward and one in the corridor, with the difference calculated by the system). = It contains two observations, H is the observation matrix. Since both sensors directly measure the differential pressure, H = . It is observation noise, assumed to follow a normal distribution with a mean of 0 and a covariance of R, i.e. R is a 2x2 diagonal matrix, and the values ​​on the diagonal are r 11 ,r 22 These represent the measurement noise variance of the two sensors, which can be calibrated based on the sensor accuracy.

[0105] Step 2: Initialization During system startup, the filter is initialized, which includes two parts: ① Initial state estimation: (Use the mean of the first valid observation as the initial state estimate).

[0106] ②Estimate the initial value of the error covariance: (Initial uncertainty is set as observation noise level).

[0107] Step 3: Real-time recursive filtering (prediction-update loop) For each new sampling time k (e.g., once per second), perform the following standard Kalman filter recursive steps: ①State prediction (time update): Based on the optimal estimate from the previous time step, predict the state at the current time step: ; Prediction error covariance: .

[0108] ②Observation update (measurement update): Calculate Kalman gain : The Kalman gain determines whether the predicted value is trusted more in this update. (small) or more trust the new observations ( big).

[0109] Obtain the actual observation vector at the current time. (Raw readings from two sensors).

[0110] Optimal estimate update of computational state: The prediction is corrected by multiplying the observation residual (the difference between the new observation and the predicted observation) by the gain.

[0111] Update the estimated error covariance: After this observation update, the uncertainty of the state estimate has been reduced.

[0112] Step 4: Output and Driver ① Update the optimal state estimate obtained at each step As a real-time output of "pressure difference in the merged ward".

[0113] ② This highly reliable pressure difference value, after random noise has been filtered out, is bound to the corresponding "ward" spatial component attribute in the BIM digital twin model through the "preset mapping relationship" described in step S2, in order to drive the three-dimensional visualization display.

[0114] ③At the same time, this fusion value also serves as one of the key boundary conditions for determining whether an event is triggered in step S3 and for CFD simulation in step S4.

[0115] The above Kalman filter fusion achieves three main effects: ① Noise reduction and smoothing: It effectively filters out random measurement noise from individual sensors, resulting in a smoother output curve; ② Improved reliability: Even if a sensor exhibits a momentary outlier, the fusion algorithm can automatically reduce the weight of unreliable observations through gain, preventing erroneous data from directly driving the system and improving data robustness, due to the presence of observations from another sensor; ③ State estimation: It provides a model- and historical "best estimate" of the true state, superior to any single sensor reading.

[0116] S3 Implementation: Configure core safety constraints (such as differential pressure range and directional gradient) in the platform's rule engine. Simultaneously, based on the multi-objective optimization method defined in S3's "Invention Content," configure a complete multi-objective optimization function and all its parameters in the platform. Under the premise of satisfying all core safety constraints, minimize the comprehensive cost function F(X) = w1*E'(X) + w2*τ'_avg(X) + w3*(1-η'(X)), which consists of energy consumption, air age, and pollutant removal efficiency.

[0117] The parameters to be implemented include , , , , , , Extreme parameters, primarily obtained through offline "pre-computation and statistical analysis," are used to normalize the sub-objectives and are a prerequisite for calculating F(X). Their specific acquisition paths are as follows: ① Sample Library Source: Before system deployment, large-scale offline operating condition simulations are conducted using the "Fast CFD Simulation Engine" construction technology described in Method S4. Specifically, using high-fidelity CFD software, based on the BIM geometric model of the target ward, hundreds or thousands of steady-state CFD simulations are performed within a pre-defined, broad parameter space (covering all possible fan frequency combinations, damper openings, typical personnel numbers and locations, door states, etc.), forming a pre-calculated sample library. Each simulation outputs the total energy consumption E and the average air age in the patient's breathing zone under that operating condition. Pollutant concentration at exhaust vents Waiting for the results.

[0118] ② Determination of extreme values: , The maximum and minimum values ​​of total energy consumption E are directly extracted from all samples in the pre-computed sample library. These values ​​represent the energy consumption boundaries that may occur under the constraints of the physical system. , Similarly, extract all samples from the sample library. The maximum and minimum values. , First, based on each sample and preset Calculate η(X) = 1 - / Then, extract the maximum and minimum values ​​from all the calculation results. Determination: This is a reference value for the steady-state emission rate of pollutants. In practical implementation, this is a preset engineering parameter. Its determination can be based on the reference values ​​for pathogen emission rates in relevant infection control standards, or a conservative estimate calculated through literature review and respiratory models of typical patients (such as vital capacity and respiratory rate). For example, it can be set as "the equivalent mass of pollutants generated by a single patient per unit time," when there are N people indoors. Scale by a factor of N. This parameter is set during system configuration and used as a known constant in optimization calculations.

[0119] The weighting coefficients (w1, w2, w3) reflect the relative importance that managers place on the three objectives of energy consumption, air quality, and sensor control efficiency. These coefficients are not uniquely determined through theoretical calculations but are set through expert evaluation and strategy formulation. An operational implementation method is as follows: ① Initial Settings: During system deployment, the hospital's infection control department, logistics management department, and design unit jointly agree on and assign initial weights. For example, under the principle of prioritizing infection control and safety, the initial settings can be: w1=0.4 (energy consumption), w2=0.3 (air age / comfort), w3=0.3 (pollutant removal efficiency), and satisfying w1 + w2 + w3 = 1. This means that during optimization, infection control and comfort objectives account for 60% of the total, while energy consumption accounts for 40%.

[0120] ② Configurability: Weights should be configurable parameters in the platform management interface. Maintenance personnel can fine-tune the weights within a certain authorized range, based on the operational phase of different wards (e.g., routine operation, peak epidemic period), season, or the hospital's overall energy-saving policy. The platform will record the weight change history.

[0121] The collaborative workflow during implementation is as follows: ① Optimization Algorithm (Online): Responsible for intelligently searching for candidate solutions X_k in the decision variable space (combination of fan frequency and valve opening). For example, using the particle swarm optimization algorithm, a group of "particles" (i.e., candidate solutions) is initialized, each particle having its own position (representing a set of control parameters) and velocity.

[0122] ② Reduced-order model ROM (online): Acts as a fast performance evaluator for the optimization algorithm. When the optimization algorithm generates a new particle (candidate scheme X_k) and needs to evaluate its merits, it does not call the time-consuming CFD solver. Instead, it combines the current real-time state (sensor data from S2) with the parameters of the candidate scheme X_k, organizes them into the input vector of the ROM, and inputs this input vector into the pre-trained ROM (which encapsulates the mapping relationship from input to output E, τ_avg, C_out). The ROM outputs the prediction results for the scheme within milliseconds: E(X_k), τ_avg(X_k), C_out(X_k).

[0123] ③ Collaborative computation process: The optimization algorithm calculates the constraint satisfaction of the proposed solution (e.g., whether the pressure difference is within the range of -15Pa to -30Pa) and the normalized sub-objective values ​​E'(X_k), τ'_avg(X_k), and η'(X_k) based on the ROM-predicted E, τ_avg, and C_out. For solutions that satisfy all constraints, its comprehensive score F(X_k) is calculated. The optimization algorithm updates the position and velocity of the particle swarm (i.e., explores new candidate solutions) based on the F(X_k) values ​​of all current particles and begins the next iteration. This process is repeated until a preset number of iterations (e.g., 20 generations) is reached or the optimal solution that meets the accuracy requirements is found. Finally, the solution X that minimizes F(X) is output. * As the optimal control instruction.

[0124] S4 and S5 Implementation: During system operation, at a certain moment, the system detected that the ward door was accidentally opened for more than 10 seconds due to the delivery of supplies. At the same time, the differential pressure sensor in the buffer room showed that the pressure rapidly rose to -2Pa, indicating that the pressure gradient was at risk of failure. The platform immediately triggered the intelligent control process: Using the real-time readings of all sensors (pressure difference, air volume, temperature, humidity, etc.) as boundary conditions, a rapid CFD simulation is performed in the digital twin model to simulate and extrapolate the preset emergency plan (fixed rule part) and online intelligent search (optimization algorithm part). The specific implementation is as follows: (1) Pre-set emergency plan (fixed rules section) During system deployment or debugging, engineers or domain experts pre-define several contingency plans for high-risk, high-certainty events in the platform's rule base, based on historical experience, design specifications, and risk assessments. These contingency plans are typically validated, conservative strategies capable of quickly restoring a basic safe state. This implementation employs the following three contingency plans: Contingency Plan A (Single-Point Forced Exhaust): {Ward exhaust fan frequency: +5 Hz, buffer room damper: remain unchanged}. Implementation logic: Rapidly increasing the exhaust of contaminants is the most direct pressure restoration method, but it may consume more energy.

[0125] Contingency Plan B (Coordination and Control): {Ward exhaust fan frequency: +3 Hz, buffer room air supply valve opening: -10%}. Implementation logic: While increasing exhaust air, slightly reduce buffer room air supply to reconstruct the pressure gradient in a more refined and energy-efficient manner.

[0126] Contingency Plan C (Backup Activation): {Ward exhaust fan frequency: remain unchanged; backup exhaust fan: start, frequency set to 35 Hz}. Implementation logic: When the main equipment may reach its limit or redundancy is required, the backup system is activated.

[0127] These contingency plans serve as an "initial population" or "safety net" for the optimization search, ensuring that there are reliable solutions available even if the optimization algorithm fails or times out.

[0128] (2) Online intelligent search (optimization algorithm part) While contingency plans are fundamental, online intelligent search is necessary to find the optimal solution in terms of overall performance (F(X)) under specific real-time conditions. This embodiment employs a particle swarm optimization algorithm for supplementary searching, with the specific implementation mechanism described below in "Algorithm Parameter Settings" and "Algorithm and ROM Collaborative Workflow": 1) Algorithm parameter settings: ① Population size (N): This refers to the number of candidate solutions maintained in each generation. Based on the document's real-time requirement of "keeping each simulation within 5 to 10 solutions," N = 6 can be set. This means that in addition to the 3 preset solutions, the optimization algorithm will generate 3 additional new solutions per round.

[0129] ② Number of iterations (Max_iter): The maximum number of iterations (rounds) the algorithm can run. To ensure a response time within seconds, Max_iter can be set to 4. That is, the algorithm will perform a maximum of 4 rounds of iterative optimization.

[0130] ③ Decision Variables and Boundaries: Define the position vector X for each particle (scheme). For example, X = [f_exhaust, f_supply, valve_buffer], representing the frequency of the ward exhaust fan, the frequency of the buffer room supply fan, and the opening of the buffer room supply valve, respectively. Each variable has its physical upper and lower limits (e.g., fan frequency 20-50 Hz, valve opening 20%-100%).

[0131] ④ Particle velocity range: Limits the maximum step size of particle updates to prevent search oscillations, for example, set to 10% of the variable range.

[0132] ⑤ Learning factors (c1, c2): Usually, the classical values ​​c1 = c2 = 2.0 are taken, which respectively control the weights of the particle learning towards its own historical best position and the group's historical best position.

[0133] ⑥ Inertia weight (w): A linear decreasing strategy can be adopted, with an initial value of w_start = 0.9 and a final value of w_end = 0.4, to strengthen global exploration in the early stage and local fine search in the later stage.

[0134] 2) Collaborative workflow between the algorithm and ROM: ① Initialization: When an event is triggered, an optimization round begins. First, the three preset plans (A, B, C) are decoded into particle positions and used as part of the initial population. Then, the remaining (6-3=3) particles are randomly generated in the decision variable space to form the complete initial population (6 particles in total).

[0135] ② Evaluation: For each particle (scheme) in the population, a reduced-order model is invoked for fast CFD simulation. The ROM receives the control parameters (X_k) and current real-time state of the particle, and outputs the predicted E(X_k), τ_avg(X_k), and C_out(X_k) within milliseconds.

[0136] ③ Calculate fitness: Based on the ROM output, calculate whether each solution satisfies all hard constraints (such as pressure difference range). For feasible solutions, calculate their comprehensive score F(X_k) according to the function defined in S3. The smaller the F(X) value, the better the fitness. Record the individual historical best position (pbest) of each particle and the global historical best position (gbest) of the entire population.

[0137] Iterative updates: For each particle, calculate its next-generation velocity and position according to the standard particle swarm update formula: ④ Speed ​​update: v_i(t+1) = w * v_i(t) + c1 * rand() * (pbest_i - x_i(t)) + c2 * rand() * (gbest - x_i(t)) Position update: x_i(t+1) = x_i(t) + v_i(t+1) Check if the new location exceeds the boundary and correct it.

[0138] ⑤ Termination and Output: Repeat steps ②-④ until the preset maximum number of iterations (4 generations) is reached. Finally, from all schemes evaluated during all iterations (including preset plans and algorithm-generated schemes), the scheme with the smallest comprehensive score F(X) is selected as the optimal control command output.

[0139] In the aforementioned embodiment, when the event "the ward door is open for more than 10 seconds and the pressure difference in the buffer room rises to -2Pa" is triggered, the system uses real-time data as boundary conditions to perform rapid CFD simulations on three preset contingency plans (A, B, and C) using a reduced-order model. The complete simulation prediction results for each contingency plan should include the following quantitative indicators to support subsequent multi-objective optimization decisions:

[0140]

[0141] Analysis of simulation prediction results and decision-making basis: ① Safety (Pressure Gradient Recovery): All three contingency plans meet the core safety constraints, but their recovery speeds differ. Contingency plan B is the fastest (3 seconds), followed by contingency plan A (5 seconds), and contingency plan C is the slowest (8 seconds). This indicates that contingency plan B performs best in rapidly eliminating infection control risks.

[0142] ② Energy efficiency (total energy consumption increase): Plan B has the least increase in energy consumption (85 W), followed by Plan A (120 W), while Plan C has the largest increase in energy consumption (150 W) due to the activation of the backup high-power fan. This indicates that Plan B is the most economical.

[0143] ③ Overall performance (multi-objective optimization function F(X)): According to the optimization function defined in S3, the comprehensive score of plan B is F(B) = 0.459, which is the lowest among the three.

[0144] The score for contingency plan A is F(A) = 0.59, and the score for contingency plan C is F(C) = 0.716.

[0145] According to the optimization objective of "minimizing the comprehensive cost function F(X)", the solution with the minimum F(X) value is the optimal solution.

[0146] The embodiments described above demonstrate that the simulation prediction results not only show that Plan B "can recover the pressure gradient within 3 seconds," but also comprehensively reveal its advantages in terms of energy consumption, air quality (air age), and pollutant removal efficiency through quantitative comparison. Specifically, it achieves the fastest recovery speed while minimizing energy consumption increase and obtaining the best comprehensive score F(X). Therefore, the platform selects Plan B as the optimal decision based on complete and quantitative multi-objective evaluation results, rather than judgment based on a single indicator, fully demonstrating the accuracy and superiority of the intelligent decision-making of this invention.

[0147] The platform sends a set of control commands to the building automation system via the BACnet protocol. The building automation system drives the frequency converter of the ward exhaust fan to increase the frequency, while simultaneously reducing the opening of the electric regulating valve on the air supply duct of the buffer room.

[0148] S6 Implementation: On the platform's command center screen, the ward area in the 3D BIM model automatically highlights and flashes a red alarm. A detailed record of the event pops up in the sidebar, including the triggering cause, comparative data of various contingency plans in the simulation, details of the executed plan, and real-time feedback curves. Management personnel can intuitively see the airflow animation showing that pollutants are effectively contained within the ward. After the event, the platform can generate a standardized report containing all process data, curves, and analysis conclusions with a single click.

[0149] Experimental data: The method described in this invention (dynamic adaptive control) improves the reliability and robustness of infection control compared to traditional static control strategies. This can be demonstrated through comparative simulation experiments, which present the following types of data: ① Comparative data on pressure gradient maintenance capabilities Key metrics: Percentage of time the pressure difference between the ward and the corridor remained within the safe range (-15Pa to -30Pa), maximum deviation, and recovery time.

[0150] Simulation scenario: Simulates a typical dynamic disturbance such as "the ward door is accidentally opened for 30 seconds".

[0151] Data example:

[0152] As can be seen from the comparison table above, the method described in this invention has a very short time when the pressure difference deviates from the safe range under door opening disturbance, and can quickly and automatically recover, increasing the compliance time ratio from 16.7% to 93.3%, which significantly proves the reliability of the system in maintaining core sensing and control constraints under dynamic disturbances.

[0153] ② Robustness comparison data of pollutant removal efficiency Key indicators: pollutant removal efficiency η at the exhaust vent, or peak pollutant concentration in the patient's breathing zone.

[0154] Simulation scenarios: Simulations were conducted under different numbers of people (1 person and 3 people, representing different pollutant release source intensities) and different air supply volumes.

[0155] Data example:

[0156] As shown in the comparison table above, the traditional fixed airflow strategy experiences a significant drop in removal efficiency and a surge in concentration when the number of personnel increases. In contrast, the method of this invention dynamically optimizes the airflow based on the real-time number of personnel (sensed via S2), maintaining a high and stable removal efficiency (approximately 0.90) under different operating conditions. This effectively controls the concentration in the breathing zone, demonstrating the system's robustness in responding to changes in source strength.

[0157] ③ System stability data under multiple disturbances and concurrency Key indicators: System control success rate, number of pressure oscillations.

[0158] Simulation scenario: Continuously simulate a series of complex disturbances: "door opening" → "increase in personnel" → "increase in filter resistance".

[0159] Data example:

[0160] As can be seen from the comparison table above, under complex and continuous disturbances, the method of this invention, through its "sensing-simulation-decision" closed loop, achieves a 100% control success rate, and the system remains stable with minimal oscillations. This directly demonstrates its high robustness in dealing with complex and uncertain operating environments.

[0161] The method described in this invention (predictive optimization decision-making based on CFD simulation) improves accuracy, foresight, and efficiency compared to traditional experience-based control or direct control without simulation. This can be demonstrated through the design of comparative simulation experiments, presenting the following types of data: ① Comparative data on the precision of regulation Key indicators: After the control measures are implemented, the deviation between the predicted and actual values ​​of key parameters (such as target pressure difference) and the target achievement rate.

[0162] Simulation scenario: Comparison of the performance of "CFD pre-simulation optimization of this invention" and "direct control based on empirical rules" in dealing with the "door opening" disturbance.

[0163] Data example:

[0164] As can be seen from the comparison table above, traditional empirical control methods, due to the lack of prediction of complex flow field coupling effects, result in actual effects that deviate significantly from the target (over-adjustment). In contrast, the method of this invention, through CFD simulation, can accurately predict the effect of the control scheme. The actual execution results are highly consistent with the predicted values, with small deviations, significantly improving the accuracy of control.

[0165] ② Comparative data for forward-looking decision-making Key metrics: The ability to mitigate risks or optimize operations, which is reflected by comparing the number of substandard solutions eliminated in pre-operation simulations with the incidence of adverse consequences afterward.

[0166] Simulation scenario: Simulates an optimization task that requires reducing energy consumption while maintaining differential pressure.

[0167] Data example:

[0168] As the comparison table above shows, traditional methods cannot predict the outcome before action, which may lead to serious consequences. The method of this invention can proactively evaluate multiple options in virtual space, eliminate risky or inefficient options in advance, and ensure that the actual implemented solution is the verified optimal solution, realizing a shift from "post-event remediation" to "pre-event prevention".

[0169] ③ Comparative data on decision-making and execution efficiency Key metrics: Time from event occurrence to generating the optimal decision (decision efficiency), and time from executing instructions to the system returning to stability (execution efficiency).

[0170] Simulation scenario: Handling the combined disturbance of "door opening + increase in personnel".

[0171] Data example:

[0172] As can be seen from the comparison table above, the method of this invention, through online simulation and deduction at the second level (efficiency), quickly completes the comparison and optimization decision of multiple options, and combined with precise control commands, achieves a significant reduction in the total processing time. Its efficiency is far higher than that of manual methods, and it is superior to simple traditional automatic control in terms of decision quality and system stability, achieving a balance between accuracy and efficiency.

[0173] Through this embodiment, the present invention successfully realizes intelligent and dynamic closed-loop control of airflow organization in negative pressure wards, transforming the passive guarantee of infection control safety into active prediction and precise control, effectively improving the hospital's ability to cope with the risk of nosocomial infection, and promoting the refined management of energy, which has significant practical value and promotion significance.

Claims

1. A BIM-based dynamic airflow control method for negative pressure wards in green and smart hospitals, characterized in that, Includes the following steps: Step S1: Construct the basic BIM model, including: creating a BIM model of the target negative pressure ward area, exporting the BIM model to IFC format and importing it into the digital twin platform, associating IoT sensor data with BIM model components through data mapping rules to form the initial state of the digital twin of the negative pressure ward. Step S2: Deploy an IoT sensing network and achieve real-time data fusion, including: deploying a multi-source sensor network within the physical space of the negative pressure ward to collect dynamic operational data in real time; uploading the dynamic data to the digital twin platform via the MQTT protocol; and performing timestamp alignment, spatial coordinate matching, multi-source state estimation, anomaly identification, and data repair on the dynamic data; and synchronously mapping and driving the digital twin with the processed real-time data; wherein the dynamic data includes environmental parameters, pressure parameters, equipment operating status, and personnel and behavioral events; Step S3: Set dynamic control rules and multi-objective optimization functions, including: Based on the sensing control specifications and green building energy-saving targets, define dynamic control rules using the event-condition-action paradigm, with maintaining the target pressure gradient as the core constraint, and with minimizing the total energy consumption of the ventilation system, minimizing the air age in key indoor areas, and maximizing the efficiency of pollutant dilution and removal as optimization sub-objectives, and establish a quantifiable multi-objective optimization function; Step S4: CFD simulation-based control strategy pre-simulation and intelligent decision-making, including: when the system detects the occurrence of dynamic triggering events or the continuous deviation of key differential pressure parameters from the safety threshold, the fast computational fluid dynamics simulation engine is automatically invoked, using the real-time sensor data processed in step S2 as the initial and boundary conditions of the simulation, and several candidate control schemes are simulated and deduced in the digital twin model at the second level, and the predicted results of total energy consumption, air age, and pollutant dilution and removal efficiency of each scheme are output; Step S5: Execute the optimal control command and implement feedback correction, including: comparing and analyzing the simulation results of each pre-simulation scheme in step S4, selecting the scheme with the best comprehensive performance according to the multi-objective optimization function, generating a structured control command set, sending it to the building automation control system to drive the actuator through the BACnet protocol, monitoring the system response in real time through the sensor network in step S2 and sending the feedback data back to the digital twin, and automatically triggering a new round of simulation and control optimization if the feedback effect does not meet expectations. Step S6: 3D visualization monitoring, early warning and data analysis, including: real-time rendering and display of pressure gradient cloud map, dynamic airflow streamline, pollutant diffusion simulation animation and equipment operation status in the BIM-based 3D visualization monitoring interface; when an anomaly is detected or a risk of infection and control is predicted, an audible and visual alarm is activated and alarm information is pushed to the mobile terminal; the entire life cycle operation data is recorded synchronously; and energy efficiency analysis reports and event traceability reports are supported.

2. The method according to claim 1, characterized in that, The real-time data fusion in step S2 includes the following steps: (1) Timestamp alignment: All sensor data are timestamped with a unified time based on the network time protocol, and data with different sampling frequencies are aligned to a unified time reference by using a sliding time window and resampling technology; (2) Spatial coordinate matching: Through the pre-calibrated spatial coordinate transformation matrix, the physical installation coordinates of all sensors and the real-time positioning coordinates of personnel are uniformly transformed into the global coordinate system of the BIM model; (3) Multi-source state estimation: For key state variables, the Kalman filter algorithm or its variant is used to fuse the readings of multiple sensors of the same or different types, filter out measurement noise, and obtain the optimal estimate. (4) Anomaly identification and data repair: The rule engine combined with statistical methods is used to identify abnormal data. For instantaneous anomalies, the filtering algorithm is used to smooth them. For multiple short-term missing data, interpolation or prediction models are used to fill them. For long-term faults, the equipment alarm is triggered and the device is switched to the preset safety mode.

3. The method according to claim 1, characterized in that, In step S2, the collection of personnel and behavioral events adopts a fusion positioning scheme that combines ultra-wideband and video recognition. Ultra-wideband provides absolute coordinates, while video recognition provides visual context information. The data are fused using a Kalman filter algorithm to obtain the number of people, their real-time location, and their movement trajectory.

4. The method according to claim 1, characterized in that, The core constraints in step S3 include the following quantitative constraints: The pressure difference between the ward and the external corridor is maintained within the range of -15Pa to -30Pa, and the directional pressure gradient of ward pressure < buffer room pressure < external corridor pressure is satisfied. The air exchange rate meets the specifications for the corresponding negative pressure ward or contaminated area; Noise level ≤ 40 dB(A); The airflow velocity at the air outlet is ≤0.2m / s; Temperature range: 22-26℃; relative humidity range: 40-60%. A replacement alarm is triggered when the pressure difference across the HEPA filter reaches twice the initial resistance.

5. The method according to claim 1, characterized in that, The dynamic control rules in step S3 are defined in the form of an event-condition-action paradigm. Each rule includes an event identifier, a trigger event type, additional logical conditions, and an execution action. The trigger event types include ward doors being continuously open for more than a threshold, the number of people in the room exceeding a threshold, and the pressure difference in the buffer room decreasing for more than a threshold. The thresholds are configurable parameters that can be adjusted by maintenance personnel through the management interface.

6. The method according to claim 1, characterized in that, The multi-objective optimization function in step S3 is constructed using the weighted sum method, and the objective function is expressed as: F(X) = w1 * E'(X) + w2 * τ' avg (X) + w3 * (1 - η'(X)); Where X is a decision vector consisting of the frequency of the supply and exhaust fans and the opening degree of the dampers; E'(X) represents the normalized total energy consumption of the ventilation system: ； = Σ (a i * f i 3 ); in, and These represent the estimated maximum and minimum total energy consumption among all feasible control schemes under the current constraints; a i f represents the coefficient related to the characteristics of the wind turbine. i Indicates the frequency of the fan; τ' avg (X) represents the normalized air age of the patient's respiratory zone: ； ； in, Indicates the local air age at the monitoring point; , These are the estimated maximum and minimum values ​​of the corresponding parameter across all feasible options; η'(X) represents the normalized pollutant dilution efficiency: ； ； in, This represents the steady-state dissipation rate of pollutants within the ward. This represents the steady-state concentration of pollutants at the exhaust vent under control scheme X. w1, w2, and w3 represent weighting coefficients, and w1 + w2 + w3 = 1.

7. The method according to claim 1, characterized in that, The fast computational fluid dynamics simulation engine in step S4 uses a reduced-order model technique to achieve simulation and deduction at the second level. Specifically, it includes: in the offline stage, high-fidelity CFD software is used to simulate preset working conditions and build a pre-calculation sample library; in the online stage, interpolation algorithms or machine learning models are used to quickly predict the simulation results of new schemes, with a single prediction time ranging from milliseconds to seconds.

8. The method according to claim 1, characterized in that, The generation of candidate control schemes in step S4 adopts a hybrid strategy that combines preset empirical rules with online intelligent search: several candidate schemes are predefined for high-risk events, and particle swarm optimization algorithm or genetic algorithm is used to search for better combination of control parameters in the constraint space; the upper limit of the number of schemes is configured according to computing resources and real-time requirements, and the number of schemes in a single simulation is controlled within 5 to 10.

9. The method according to claim 1, characterized in that, The three-dimensional visualization monitoring in step S6 is implemented based on the WebGL technology stack. The pressure gradient cloud map is generated by the Kriging interpolation algorithm and superimposed on the indoor surface of the room with a semi-transparent texture. The dynamic airflow streamline is generated by the fourth-order Runge-Kutta method and displayed as a tubular model with directional arrows. The pollutant diffusion simulation is implemented using a particle system.

10. A BIM-based green smart hospital negative pressure ward dynamic airflow control system, used to implement the method described in any one of claims 1-9, characterized in that, include: The BIM digital twin module is used to build and store the BIM model of the target negative pressure ward area, forming a digital twin of the negative pressure ward. The Internet of Things (IoT) sensing module includes a multi-source sensor network deployed in the physical space of the negative pressure ward, used to collect dynamic operating data in real time, and to perform timestamp alignment, spatial coordinate matching, multi-source state estimation, anomaly identification, and data repair on the data. The rule engine and optimization module are used to define dynamic control rules using the event-condition-action paradigm, and to establish a multi-objective optimization function with maintaining the target pressure gradient as the core constraint and minimizing the total energy consumption of the ventilation system, the shortest air age, and the highest pollutant dilution efficiency as optimization sub-objectives. The CFD simulation and pre-simulation module is used to simulate and extrapolate candidate control schemes in seconds using real-time sensor data as initial and boundary conditions when a dynamic triggering event or pressure difference abnormality is detected, and outputs the prediction results of each scheme. The instruction execution and feedback module is used to generate a set of structured control instructions, which are sent to the building automation system via the BACnet protocol to drive the actuators, and closed-loop correction is performed using sensor feedback data. The 3D visualization monitoring module is used to render and display pressure gradient cloud maps, dynamic airflow streamlines, pollutant diffusion simulation animations and equipment operating status in a BIM-based 3D interface in real time, and trigger early warning push when abnormalities occur. The data storage and analysis module is used to synchronously record operational data throughout the entire lifecycle and supports the generation of energy efficiency analysis reports and event tracing reports.

Images