A digital twin-driven odor source tracing control and simulation exercise system and method

Abstract

This invention belongs to the field of environmental monitoring and intelligent management technology, specifically relating to a digital twin-driven odor source tracing control and simulation exercise system and method. The system includes a physical space layer, a data interconnection and spatiotemporal synchronization layer, a digital twin construction layer, a hybrid intelligent engine layer, an autonomous control layer, and a digital sandbox simulation exercise module. Through a multi-layered collaborative architecture, the system constructs a complete closed-loop path from data acquisition, virtual-real mapping, intelligent analysis to autonomous control and simulation exercise. This allows high-precision source tracing, autonomous closed-loop control, and virtual-real feedback simulation exercise to coexist within the same technical framework, thus solving the problems of functional fragmentation, delayed response, and insufficient emergency preparedness in traditional systems, and achieving integrated intelligent management and control of odor pollution incidents.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring and intelligent control technology, specifically relating to a digital twin-driven odor source tracing control and simulation exercise system and method. Background Technology

[0002] Odor pollution, especially odor problems caused by pollutants such as VOCs, H2S, and NH3, has become a challenge for environmental supervision in chemical and industrial parks due to its hidden sources, complex diffusion influenced by meteorological and topographical factors, and low human perception threshold. Traditional supervision methods mainly rely on discrete online monitoring points and manual inspections, which have revealed a series of shortcomings in practice.

[0003] First, existing systems generally suffer from a disconnect between the virtual and the real. Monitoring data is mostly presented in two-dimensional charts or simple heat maps, failing to intuitively and dynamically display the diffusion trajectory and spatial distribution of odorous smoke in three-dimensional space. Source tracing relies heavily on subjective judgment based on the experience of operators, resulting in low efficiency and poor positioning accuracy, especially in complex leakage scenarios with multiple overlapping sources, making it difficult to quickly identify the main responsible source.

[0004] Secondly, traditional methods for tracing the source of pollution are rather crude. Most systems can only trigger concentration threshold alarms and require manual confirmation based on video images, lacking precise tracing mechanisms such as inverse trajectory retrieval based on atmospheric physical models and spatial probability quantification. This makes it difficult for the system to effectively distinguish the contributions of different companies in practical applications, thus failing to meet the needs of precise law enforcement.

[0005] Secondly, existing systems lack autonomous control capabilities. Current mainstream technologies still rely on a passive "monitoring-simulation-alarm" model, ultimately requiring manual dispatching of orders, manual start-up and shutdown of treatment equipment (such as sprayers and fans), and acceptance of results. This model has a slow response time, often with a lag of tens of minutes or even hours from the occurrence of a leak to the initiation of control measures, missing the optimal opportunity to suppress the spread of pollution, and making it difficult to quantify and optimize the treatment effect in a closed-loop manner.

[0006] Finally, existing technical solutions lack simulation and drill capabilities. Emergency preparedness in industrial parks primarily relies on periodic on-site drills, which are not only costly and difficult to organize, but also cannot cover all possible accident scenarios (such as multi-point leaks under extreme weather conditions). Managers cannot conduct low-cost, high-frequency rehearsals and assessments of emergency plans in a virtual environment, thus limiting their practical emergency response capabilities.

[0007] In recent years, digital twin technology has begun to be applied in some industrial parks, such as the Hangzhou Bay Shangyu Economic Development Zone and the Zhenhai Petrochemical Park, but its application depth is still insufficient. These systems mainly solve the problems of visualization and data integration. Their traceability modules are still based on traditional trajectory simulation or fingerprint database matching, and there are no reports of using advanced algorithms such as graph attention networks for probabilistic and accurate positioning. Their prediction modules mostly use single-mechanism models, which are slow to calculate and difficult to achieve ultra-real-time simulation. At the control level, most of them only achieve linkage start and stop with some devices and have not formed a closed loop of the entire process of autonomous generation, distribution, evaluation and optimization of policies based on reinforcement learning. In addition, the published patent applications (such as CN121365615A) are mainly aimed at closed or semi-closed spaces. Their technical architecture and models are not suitable for complex diffusion scenarios in open atmospheres, and they also lack autonomous control and simulation exercise functions. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, this invention provides a digital twin-driven odor source tracing control and simulation exercise system and method. This solves the problems of existing odor monitoring systems being unable to simultaneously achieve high-precision source tracing in open spaces, autonomous closed-loop control without human intervention, and virtual-real feedback simulation exercises under the same technical architecture, resulting in delayed response, low governance efficiency, and insufficient emergency preparedness.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A digital twin-driven odor source tracing control and simulation exercise system and method are provided, including a physical space layer, a data interconnection and spatiotemporal synchronization layer, a digital twin construction layer, a hybrid intelligent engine layer, an autonomous control layer, and a digital sandbox simulation exercise module. The physical space layer includes a multimodal sensing network and executable control devices, used to collect raw data including concentration, meteorology, and equipment status; The data interconnection and spatiotemporal synchronization layer has a two-way communication connection with the physical space layer. It is used to receive raw data, clean and spatiotemporally align the raw data to obtain standardized data, transmit the standardized data outward, and receive control commands and send them to the physical space layer. The digital twin construction layer and the data interconnection and spatiotemporal synchronization layer have a one-way communication connection, which is used to construct a 3D scene based on geographic information, building information model and laser point cloud, and integrate the mechanism model and data-driven proxy model to output real-time concentration field, meteorological field and correction coefficient; The hybrid intelligent engine layer and the digital twin construction layer have a one-way communication connection. This is used to output the source tracing results through a two-stage mechanism of first using Lagrange reverse trajectory coarse localization and then using multi-head graph attention network correction. The system also outputs diffusion inference results by adaptively calling the proxy model or mechanism model according to the prediction duration and meteorological conditions, and by periodically correcting the rules of the mechanism model. The autonomous control layer is bidirectionally connected to the hybrid intelligent engine layer, the digital twin construction layer, and the data interconnection and spatiotemporal synchronization layer. It is used to generate control strategies through reinforcement learning by taking the source tracing results, diffusion inference results, and the real-time status output by the digital twin construction layer as inputs, and to issue control commands to the data interconnection and spatiotemporal synchronization layer according to priority. It also evaluates and adaptively optimizes the strategies based on the governance effect. The digital sand table simulation exercise module has a two-way communication connection with the digital twin construction layer and the hybrid intelligent engine layer. It is used to build virtual leakage scenarios, rehearse emergency strategies, and conduct two-way interaction of virtual and real data with the digital twin construction layer and the hybrid intelligent engine layer.

[0010] The beneficial effects of adopting the above technical solution are as follows: This digital twin-driven odor source tracing, control, and simulation system constructs a complete closed-loop path from data acquisition, virtual-real mapping, intelligent analysis to autonomous control and simulation through a multi-layered collaborative architecture consisting of a physical space layer, a data interconnection and spatiotemporal synchronization layer, a digital twin construction layer, a hybrid intelligent engine layer, an autonomous control layer, and a digital sandbox simulation module. This allows high-precision source tracing, autonomous closed-loop control, and virtual-real feedback simulation to coexist within the same technical framework, thereby solving the problems of functional fragmentation, delayed response, and insufficient emergency preparedness in traditional systems, and achieving integrated intelligent management and control of odor pollution incidents. Specifically, the physical space layer includes a multimodal sensing network and executable control devices, which provide basic raw data sources and final control execution capabilities. The concentration, meteorological, and equipment status data collected by the multimodal sensing network are the inputs for all subsequent analysis and decision-making, while the executable control devices ensure that the control commands generated by the upper layer can be implemented, forming a complete data and control loop from the physical world to the digital world. The data interconnection and spatiotemporal synchronization layer, through bidirectional communication with the physical space layer, cleans, aligns, and standardizes the collected raw data, ensuring the data quality and spatiotemporal consistency of the data entering the digital twin construction layer. Simultaneously, it accurately transmits control commands issued by the autonomous control layer to the execution devices in the physical space layer, guaranteeing the real-time and reliable information transmission between the virtual and real worlds. The digital twin construction layer, through unidirectional communication with the data interconnection and spatiotemporal synchronization layer, can utilize geographic information, building information modeling (BIM), and laser point clouds to construct a 3D scene reflecting the actual physical environment of the park. It integrates mechanistic models and data-driven proxy models, outputting real-time concentration fields, meteorological fields, and correction coefficients. This provides a high-fidelity virtual operating environment synchronized with the physical world in real-time for upper-level tracing, deduction, and control, ensuring that all subsequent intelligent decisions are made based on an accurate digital mirror. The hybrid intelligent engine layer, through a one-way communication connection with the digital twin construction layer, achieves accurate source tracing using a two-stage mechanism: first, coarse localization using a Lagrange inverse trajectory, followed by fine-tuning using a multi-head graph attention network. Simultaneously, it adaptively calls proxy models or mechanistic models based on prediction duration and meteorological conditions, with the mechanistic model periodically revising the output rules for diffusion inference results. This allows the system to significantly improve computational response speed while ensuring source tracing accuracy and the physical rationality of the inference. The autonomous control layer, through bidirectional communication connections with the hybrid intelligent engine layer, the digital twin construction layer, and the data interconnection and spatiotemporal synchronization layer, takes source tracing results, inference results, and the real-time status of the digital twin as input. It automatically generates control strategies using reinforcement learning and issues instructions according to priority. Simultaneously, it evaluates and adaptively optimizes strategies based on governance effectiveness, achieving a fully automated, closed-loop process from perception, decision-making, execution to evaluation and optimization, significantly improving the timeliness and effectiveness of odor incident handling.The digital sand table simulation exercise module, through bidirectional communication with the digital twin construction layer and the hybrid intelligent engine layer, can construct virtual leakage scenarios and rehearse emergency strategies. At the same time, it can conduct bidirectional interaction of virtual and real data with the digital twin construction layer and the hybrid intelligent engine layer, enabling managers to conduct drills and strategy verification for various odor leakage scenarios in a zero-risk virtual environment. Through data interaction, it can also conduct model training and algorithm optimization, forming a virtuous cycle mechanism of virtual promoting reality and reality driving virtual.

[0011] Furthermore, the improved Gaussian plume model integrated into the digital twin construction layer introduces a terrain correction coefficient. With building shading correction factor Its expression is:

[0012] in, The pollution source is strong; Average wind speed; For lateral diffusion parameters; For vertical diffusion parameters; High efficiency of source; This is a terrain correction factor, determined by ground roughness and slope; The building shading correction factor is determined by the building's average height, surface density, and the distance from the calculation point to the building's centerline.

[0013] The beneficial effects of adopting the above technical solution are as follows: the improved Gaussian plume model introduces a terrain correction coefficient into the classical diffusion expression. With building shading correction factor This enables the model to dynamically reflect the actual impact of ground roughness, slope, and the average height, surface density, and distance of buildings on pollutant diffusion. It can more accurately depict the spatial distribution of concentration under complex underlying surface conditions in industrial parks, thereby providing a concentration field and correction parameters that are closer to the real physical process for subsequent two-stage source tracing and diffusion simulation, directly improving the reliability of odor location and impact range prediction.

[0014] Furthermore, in the hybrid intelligent engine layer, the coarse localization of the Lagrange reverse trajectory uses the fourth-order Runge-Kutta method to solve the reverse trajectory equation, and the integration step size is dynamically adjusted according to the real-time wind speed, with the overall source tracing time ≤60s; the multi-head image attention network correction uses 8 attention heads, and areas with an output probability ≥0.8 are identified as high-suspect pollution source areas.

[0015] The beneficial effects of adopting the above technical solution are as follows: By combining the coarse localization of the Lagrange inverse trajectory with the fine correction of the multi-head graph attention network, a two-stage source tracing mechanism is formed, which first performs rapid inversion and then makes fine probabilistic judgment. Specifically, the inverse trajectory solution uses the fourth-order Runge-Kutta method and dynamically adjusts the integration step size according to real-time wind speed, ensuring computational stability and solution efficiency under different meteorological conditions, thus confining the overall source tracing time to a short period. The multi-head graph attention network uses multiple attention heads to capture pollution diffusion patterns from different spatial dimensions and outputs high-suspicious areas with a high probability threshold, avoiding the bias that may be caused by a single attention head and ensuring the reliability of the localization results.

[0016] Furthermore, the autonomous control layer employs deep deterministic policy gradient reinforcement learning for decision-making, and its reward function expression is as follows:

[0017] in, For comprehensive risk scoring; Real-time concentration at sensitive points; The concentration threshold specified by ambient air quality standards; To control equipment operating costs, the accounting dimensions include equipment energy consumption, chemical consumption, and maintenance losses; The baseline cost based on historical operational data; The rate of decrease in concentration; Indicator functions.

[0018] The beneficial effects of adopting the above technical solution are as follows: the reward function quantifies and weights three indicators: sensitive point concentration control, equipment operating costs, and the achievement of treatment standards. This allows deep deterministic strategy gradient reinforcement learning to balance the dual objectives of reducing concentration and reducing costs when making decisions. The concentration term has the highest weight, reflecting the control orientation of prioritizing air quality at sensitive points; the cost term constrains equipment energy consumption and maintenance losses, avoiding over-treatment; and the treatment efficiency term provides positive incentives for achieving the set concentration reduction rate, guiding the model to optimize towards achieving the desired effect. This forms a balanced decision-making basis, enabling the reinforcement learning model to autonomously learn control strategies that meet both environmental protection requirements and economic efficiency during virtual trial and error.

[0019] Furthermore, the autonomous control layer uses the concentration decrease rate... Achieve closed-loop evaluation and optimization, concentration reduction rate The expression is:

[0020] in To control the average concentration before the instruction is issued; To control the average concentration after the command is executed; For the rate of decrease in concentration, when When the level is less than 30%, the secondary enhanced control will be automatically activated.

[0021] The beneficial effects of adopting the above technical solution are: by collecting average concentrations before and after the execution of control commands and calculating the concentration reduction rate, the system can objectively judge the actual effectiveness of the current control measures. When the concentration reduction rate... When the level falls below the set threshold, a secondary enhanced control is automatically triggered, forming a closed-loop mechanism from evaluation to optimization. This avoids the time delay of manual inspection and effect judgment, and ensures that the control intensity can be upgraded in a timely manner when the treatment effect is not good, thereby improving the reliability and response efficiency of odor pollution control.

[0022] Furthermore, the autonomous control layer also includes a comprehensive control priority scoring module, which is used to calculate the control priority score of each pollution source when multiple pollution sources occur simultaneously. Its expression is:

[0023] in, Priority scoring for comprehensive pollution source management; Risk values ​​for toxicity and source strength; The distance response value is the distance from the sensitive point. The leakage persistence index is represented by 0.6, 0.3, and 0.1, which are the weights of the three indicators: risk, sensitivity, and urgency, respectively. The autonomous control layer scores based on the aforementioned control priorities. The system automatically allocates control resources of different proportions and intensities based on the level of the system.

[0024] The beneficial effects of adopting the above technical solution are as follows: The comprehensive control priority scoring module assigns different weights to three indicators: toxicity and source strength risk, distance to sensitive points, and leakage persistence. It then quantifies and scores the control order when multiple pollution sources occur simultaneously. The autonomous control layer automatically schedules control resources of different proportions and intensities based on the scores, enabling the system to prioritize the handling of high-risk pollution sources that are close to sensitive points and continuously leaking in multi-source leakage scenarios. This avoids the lag and subjective bias of manual judgment and achieves the rational allocation and coordinated scheduling of control resources.

[0025] Furthermore, the deployment rules for the multimodal sensing network in the physical space layer are as follows: sensors are deployed 30m to 80m downwind of the pollution source, at the factory boundary, and at the inflection point of the corridor; sensors are deployed 50m to 100m upwind of sensitive points; key control areas are deployed in 50m*50m grids, and general areas are deployed in 100m*100m grids.

[0026] The beneficial effects of adopting the above technical solution are as follows: the deployment rules are based on fixed-point deployment at key locations such as downwind of pollution sources, factory boundaries, corridor inflection points, and upwind of sensitive points. Combined with different grid densities for key and general areas, a three-level spatial matching relationship is formed from pollution source capture to sensitive point protection, providing the system with spatially representative concentration monitoring data. Simultaneously, the concentration reduction rate... Through calculation and threshold judgment, the system can quantitatively evaluate the control effect and automatically activate secondary enhanced control when the effect is insufficient. In addition, by combining deployment rules and closed-loop evaluation optimization, the system ensures both the accuracy and capture capability of the sensing data and the dynamic adjustment and continuous optimization of the control strategy, thereby improving the targeting and effectiveness of odor control.

[0027] Furthermore, the bidirectional data communication rules between the digital sand table simulation exercise module and the digital twin construction layer and hybrid intelligent engine layer are as follows: the digital twin construction layer outputs real-time 3D scenes and equipment status to the digital sand table, and the hybrid intelligent engine layer outputs source tracing algorithms and diffusion models to the digital sand table; the digital sand table outputs simulated leakage parameters to the digital twin construction layer for model calibration, and outputs simulated strategy execution results to the hybrid intelligent engine layer for algorithm optimization.

[0028] The beneficial effects of adopting the above technical solution are as follows: the two-way data communication rules establish a two-way data channel between the digital sandbox simulation exercise module and the digital twin construction layer and hybrid intelligent engine layer, enabling real-time 3D scenes, equipment status, source tracing algorithms, and diffusion models to flow from the real system to the digital sandbox for constructing a high-fidelity virtual exercise environment; at the same time, the simulated leakage parameters and strategy execution results generated in the digital sandbox can be fed back to the digital twin construction layer and hybrid intelligent engine layer for model calibration and algorithm optimization. Through this two-way reuse mechanism of virtual and real data, the simulation exercise is no longer independent of the real system, but becomes an effective input for model iteration and algorithm improvement, realizing mutual feedback between the virtual and the real, and improving the overall system's adaptability and prediction accuracy to complex leakage scenarios.

[0029] Based on the aforementioned digital twin-driven odor source tracing control and simulation exercise system, this invention also provides a digital twin-driven odor source tracing control and simulation exercise method, comprising the following steps: S1. Raw Data Acquisition and Upload Raw data on odor concentration, meteorological parameters, and equipment status are collected through the physical space layer and uploaded to the data interconnection and spatiotemporal synchronization layer. S2, Data Cleaning and Spatiotemporal Alignment The data interconnection and spatiotemporal synchronization layer cleans and aligns the raw data spatiotemporally, then outputs standardized data to the digital twin construction layer. S3, Real-time Concentration Field and Meteorological Field Generation The digital twin construction layer generates real-time concentration and weather fields based on standardized data and pre-stored 3D models, and outputs them to the hybrid intelligent engine layer. S4, Two-Stage Source Tracing and Hybrid Driven Deduction First, candidate regions are obtained by inverting the Lagrange reverse trajectory with the over-standard sensor as the endpoint. Then, the probability of pollution sources is calculated on the spatial grid within the candidate regions using a multi-head graph attention network, and a probability cloud map is output. At the same time, a surrogate model or a mechanism model is selected to perform diffusion inference based on the prediction duration. S5. Autonomous control command issuance The autonomous control layer acquires the source tracing results, inference results, and real-time status of the twin, generates control strategies through reinforcement learning, and issues instructions to the executable control devices in the physical space layer according to priority via the data interconnection and spatiotemporal synchronization layer. S6. Evaluation of Governance Effectiveness and Adjustment of Strategies After the executable control device executes the command, the autonomous control layer collects the concentration data after treatment, calculates the concentration reduction rate, and evaluates the treatment effect based on the reduction rate. If the requirements are not met, the strategy is automatically adjusted or secondary enhanced control is initiated. S7, Virtual Scene Construction and Data Feedback The digital sandbox simulation module receives real-time data from the digital twin construction layer and the hybrid intelligent engine layer to build a virtual scene, and feeds back the simulated leakage parameters and strategy execution results to the digital twin construction layer and the hybrid intelligent engine layer for model training and parameter optimization.

[0030] The beneficial effects of adopting the above technical solution are as follows: This digital twin-driven odor source tracing control and simulation exercise method, through the sequential execution of raw data acquisition and uploading, raw data cleaning and alignment, concentration field and meteorological field generation, two-stage source tracing and hybrid-driven simulation, autonomous control decision-making and issuance, governance effect evaluation and strategy adjustment, and virtual scenario construction and data feedback, solidifies the data flow and functional collaboration between the various layers of the system into a complete process. This allows high-precision source tracing, ultra-real-time simulation, autonomous closed-loop control, and virtual-real feedback simulation exercises to be orderly connected under the same temporal logic, avoiding confusion in the calling between functional modules, and providing a standardized operation path for the monitoring, handling, and exercise of odor pollution events. Specifically, S1 directly collects and uploads raw data of odor concentration, meteorological parameters, and equipment status through the physical space layer, establishing the data source of the entire method. The upload of raw data provides real input for all subsequent processing steps, ensuring that the information relied upon by subsequent analysis, decision-making, and simulation all comes from the actual on-site conditions. S2 cleans and spatiotemporally aligns the raw data, eliminating sensor outliers and inconsistencies in time and spatial coordinates. The data is then standardized before transmission, ensuring that the data entering the digital twin construction layer shares the same quality benchmark and spatiotemporal reference system. This provides a reliable data foundation for 3D scene modeling and the output of concentration field computer correction coefficients. S3 generates real-time concentration and meteorological fields based on standardized data and pre-stored 3D models, inputting them into the hybrid intelligent engine layer. This achieves the fusion of static 3D models and dynamic sensing data, enabling the upper-level engine to conduct source tracing and deduction in a digital environment synchronized with the physical park in real time, avoiding misjudgments caused by model-field disconnect. S4, after triggering the exceedance time, first uses the exceedance sensor as the endpoint to invert the Lagrange reverse trajectory to obtain candidate areas. Then, a multi-head graph attention network calculates the pollution source probability of the spatial grid within the candidate areas and outputs a probability cloud map. Simultaneously, based on the prediction duration, a surrogate model or mechanism model is flexibly selected for diffusion deduction. This parallel design of two-stage source tracing and hybrid deduction allows source tracing and diffusion prediction to be completed simultaneously, shortening the overall waiting time from monitoring to obtaining prediction information. S5 obtains the source tracing results, inference results, and real-time status of the twin from the autonomous control layer. It generates control strategies through reinforcement learning and issues instructions to executable control devices according to priority through the data interconnection and spatiotemporal synchronization layer. It can connect intelligent decision-making with physical execution, realize the automatic jump from digital spatial analysis to physical spatial intervention, and complete the generation and issuance of control instructions without human intervention.After the equipment executes the command, S6 collects the concentration data after treatment, calculates the concentration reduction rate, and evaluates the treatment effect based on the concentration reduction rate. When the treatment effect does not meet the requirements, it automatically adjusts the strategy or initiates secondary enhanced control. This closed-loop evaluation mechanism enables the digital twin-driven odor source tracing control and simulation exercise method to have the ability to self-judge and actively optimize, ensuring that the treatment measures can be dynamically adjusted according to the actual effect, avoiding the continuous operation of ineffective or inefficient control strategies. S7 receives real-time data from the digital twin construction layer and the hybrid intelligent engine layer through the digital sandbox simulation exercise module to construct a virtual scene, and feeds back the simulated leakage parameters and strategy execution results to the above two layers for model training and parameter optimization. This bidirectional data flow makes the simulation no longer independent of the real system, but becomes a continuous input source for model iteration and algorithm improvement, enhancing the method's adaptability and long-term stability to unknown leakage scenarios.

[0031] In summary, the digital twin-driven odor source tracing control and simulation exercise system and method provided by this invention have the following beneficial effects: (1) The digital twin-driven odor source tracing control and simulation exercise system is composed of a multi-layer collaborative architecture consisting of a physical space layer, a data interconnection and spatiotemporal synchronization layer, a digital twin construction layer, a hybrid intelligent engine layer, an autonomous control layer and a digital sandbox simulation exercise module. Two-way or one-way communication connections are established between each layer, forming a complete closed-loop path from data acquisition, data cleaning and alignment, three-dimensional scene construction, intelligent analysis to autonomous control and simulation exercise. This enables high-precision source tracing, autonomous control without human intervention and virtual-real feedback simulation exercise to be integrated and coexist under the same technical framework. It solves the problems of functional fragmentation, delayed response and insufficient emergency preparedness of traditional odor monitoring systems, and realizes integrated intelligent management and control of odor pollution incidents in industrial parks.

[0032] (2) The hybrid intelligent engine layer in the digital twin-driven odor source tracing control and simulation exercise system adopts a two-stage time-series fusion method for source tracing, which first uses Lagrange reverse trajectory coarse positioning and then multi-head graph attention network fine correction. At the same time, it adaptively calls the surrogate model or mechanism model according to the prediction duration and meteorological conditions, and outputs the diffusion inference results by the mechanism model periodically. This allows the source tracing process to compress the overall time to less than 60 seconds while ensuring positioning accuracy (error ≤ 10m). The diffusion inference takes into account the rapid response capability of the surrogate model and the physical authenticity of the mechanism model, realizing ultra-real-time prediction on a time scale of 15 to 60 minutes, providing timely and reliable decision-making basis for the autonomous control layer.

[0033] (3) The autonomous control layer in the digital twin-driven odor source tracing control and simulation exercise system takes the source tracing results, simulation results and the real-time status output by the digital twin construction layer as input. It automatically generates control strategies through deep deterministic strategy gradient reinforcement learning, and issues instructions to the executable control devices in the physical space layer through the data interconnection and spatiotemporal synchronization layer according to the quantification priority. At the same time, it conducts real-time evaluation and strategy adaptive optimization of the treatment effect based on the concentration reduction rate, forming a closed loop of "decision-issuance-execution-evaluation-optimization" with no human intervention throughout the entire process.

[0034] (4) The digital sand table simulation exercise module in the digital twin-driven odor source tracing control and simulation exercise system can achieve bidirectional data communication with the digital twin construction layer and the hybrid intelligent engine layer, so that the simulation parameters and strategy execution results generated in the virtual exercise can be used in reverse for model calibration and algorithm optimization. This realizes the bidirectional reuse of virtual and real data, which significantly improves the system's adaptability to complex leakage scenarios and long-term stability.

[0035] (5) The digital twin-driven odor source tracing control and simulation exercise method, through the sequential execution of the following steps, including raw data acquisition and uploading, raw data cleaning and spatiotemporal alignment, real-time concentration field and meteorological field generation, dual-stage source tracing and hybrid-driven simulation, autonomous control command issuance, governance effect evaluation and strategy adjustment, virtual scene construction and data feedback, solidifies the data flow and functional collaboration between the physical space layer, data interconnection and spatiotemporal synchronization layer, digital twin construction layer, hybrid intelligent engine layer, autonomous control layer and digital sand table simulation exercise module into a standardized operation process. This ensures that high-precision source tracing, ultra-real-time simulation, autonomous closed-loop control and virtual-real mutual feedback simulation exercise are orderly connected under the same temporal logic, avoiding confusion in the calls between modules. It provides a clear and reproducible operation path for the monitoring, handling and exercise of odor pollution incidents in industrial parks, and also provides a unified framework at the method level for technology transfer and system deployment between different parks. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the structure and data flow between the components of the digital twin-driven odor source tracing control and simulation exercise system of the present invention. Figure 2 This is a flowchart of the digital twin-driven odor source tracing control and simulation exercise method of the present invention. Detailed Implementation

[0037] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0038] like Figure 1 As shown, the digital twin-driven odor source tracing control and simulation exercise system provided by the present invention includes a physical space layer, a data interconnection and spatiotemporal synchronization layer, a digital twin construction layer, a hybrid intelligent engine layer, an autonomous control layer, and a digital sandbox simulation exercise module. Through the coordinated operation of six layers of modules—the physical space layer, the data interconnection and spatiotemporal synchronization layer, the digital twin construction layer, the hybrid intelligent engine layer, the autonomous control layer, and the digital sand table simulation exercise module—a complete technical path has been constructed, from data acquisition, synchronous mapping, and intelligent analysis to autonomous control and simulation exercise. The physical space layer solves the problems of on-site perception and execution; the data interconnection and spatiotemporal synchronization layer ensures the real-time and reliable transmission of information between the virtual and real worlds; the digital twin construction layer provides a high-fidelity virtual environment synchronized with the physical park; the hybrid intelligent engine layer enables accurate source tracing and rapid simulation; the autonomous control layer completes closed-loop governance without human intervention; and the digital sand table simulation exercise module can perform pre-rehearsals and iterative optimizations in virtual scenarios. This allows high-precision source tracing, autonomous closed-loop control, and virtual-real feedback simulation exercises to coexist and integrate within the same technical framework, thereby overcoming the problems of functional fragmentation, delayed response, and insufficient emergency preparedness in traditional systems.

[0039] The physical space layer includes a multimodal sensing network and executable control devices. The multimodal sensing network includes photoionization sensors, electrochemical sensors, miniature weather stations, and inspection drones. The photoionization sensors have a measurement range of 0.01~1000 μg / m³. 3 The electrochemical sensor has detection limits of ≤0.001ppm for H2S and NH3. The micro weather station has a wind direction measurement accuracy of ±1° and a wind speed measurement accuracy of ±0.1m / s. The UAV inspection system is equipped with RTK-GPS (real-time dynamic differential positioning module) with a positioning accuracy of ±0.5m. Controllable devices include a high-pressure spray device, a variable frequency fan, and intelligent valves. The high-pressure spray device has an atomization particle size of 10~50μm and an adjustable spray angle within the range of 0~90°. The variable frequency fan operates at a frequency of 10~50Hz, and the intelligent valves have an opening adjustment range of 0~100%. The physical space layer connects to the industrial gateway via RS485 and Ethernet interfaces. The industrial gateway has dual RJ45 network ports, a fifth-generation mobile communication module, and eight RS485 interfaces, with an IP65 protection rating.

[0040] A bidirectional communication connection is established between the data interconnection and spatiotemporal synchronization layer and the physical space layer. The data interconnection and spatiotemporal synchronization layer uses MQTT (Message Queuing Telemetry Transport) protocol, Quality of Service Level 2, for data synchronization. Concentration and meteorological data are transmitted at a frequency of 1Hz, equipment status data at a frequency of 0.5Hz, and the transmission delay of control commands and early warning signals is no higher than 500ms. At the edge nodes, [the following is used]. The criteria eliminate sensor outliers, employ a moving average window (window size = 5) for data smoothing and filtering, and uniformly use millisecond-level Coordinated Universal Time timestamps and WGS84 (World Geodetic System 1984) coordinate system for spatiotemporal alignment. When the sensor detects that the concentration exceeds the threshold specified by the ambient air quality standard for three consecutive cycles (1 second per cycle), it automatically triggers the corresponding point in the digital twin to flash red at a frequency of 2Hz, generates a 50m warning circle, initiates surrounding video surveillance recording, and dispatches the nearest drone to automatically take off for inspection.

[0041] A one-way communication connection is established between the digital twin construction layer and the data interconnection and spatiotemporal synchronization layer. The digital twin construction layer is based on GIS (Geographic Information System) data, BIM (Building Information Modeling) data, and laser point cloud data (point cloud density ≥ 100 points / m²). 2 A 1:1 three-dimensional model was constructed, with a model coordinate error of no more than 0.5m. The integrated mechanistic models in this layer include an improved Gaussian plume model, a CFD model (three-dimensional convection-diffusion equation), and a pollutant photochemical reaction model. The data-driven surrogate model adopts a hybrid neural network of convolutional neural network and long short-term memory network (CNN-LSTM).

[0042] Among them, the improved Gaussian plume model integrated in the digital twin construction layer introduces a terrain correction coefficient. With building shading correction factor Its expression is:

[0043] in, The pollution source is strong; Average wind speed; For lateral diffusion parameters; These are vertical diffusion parameters, and both are determined based on the Pasquill-Gifford atmospheric stability classification (Classes A to F); High efficiency of source; This is the terrain correction factor, and its value range is... ， Determined by ground roughness (0~0.5m) and slope (0~15°); This is the building shading correction factor, with a value range of [value range missing]. It is determined by the average building height (≤50m), surface density (≤0.3), and the distance from the calculation point to the building centerline (10~50m).

[0044] The expression for the CFD model (three-dimensional convection-diffusion equation) is as follows:

[0045] in, The concentration change over time; The spatial and temporal distribution concentration of pollutants; The three-dimensional turbulent velocity vector is given by Turbulence model calculations; For convective transport; The turbulent diffusion coefficient; For turbulent diffusion; The spatiotemporal distribution of source and sink terms represents pollution source emissions and pollutant removal. Flow field calculations employ... The turbulence model is closed with Reynolds stress, and the PISO (Pressure Implicit with Splitting of Operators) algorithm is used to achieve pressure-velocity coupling solution. The boundary conditions are dynamically updated in real time by the three-dimensional terrain, building and facility information of the digital twin construction layer.

[0046] The expression for the photochemical reaction model of pollutants is:

[0047] in, The rate of change of pollutant concentration over time, expressed in mol / m³. 3 s; The instantaneous concentration of the target pollutant, in mol / m³. 3 ; This refers to the ambient ozone concentration, expressed in mol / m³. 3 ; Activation energy, expressed in J / mol; The gas constant is R = 8.314 J / (mol). K); Temperature is the thermodynamic temperature, and its unit is K. and These are the reaction orders for the pollutants and ozone, respectively. This is a pre-exponential factor, and its unit depends on the overall reaction order ( , usually (mol) m -3 ) 1-(n+m) / s. This photochemical reaction model of pollutants is used to correct for the concentration decay of pollutants due to oxidation reactions in diffusion simulations.

[0048] The input parameters of the hybrid neural network surrogate model, which combines convolutional neural networks and long short-term memory networks (CNN-LSTM), include wind speed, wind direction, temperature, humidity, pollution source intensity, spatial coordinates, and terrain correction coefficients. and building shading correction factor The output is a three-dimensional spatial concentration field. Model training data includes historical monitoring data and simulation data from a CFD model (three-dimensional convection-diffusion equation). This surrogate model's derivation speed is ≥100 times faster than the CFD model (three-dimensional convection-diffusion equation), with an average error ≤5%.

[0049] A one-way communication connection is established between the hybrid intelligent engine layer and the digital twin construction layer. The hybrid intelligent engine layer adopts a two-stage mechanism to output the source tracing results: first, coarse localization using the Lagrange inverse trajectory, and then fine correction using a multi-head graph attention network.

[0050] The first stage involves coarse localization of the Lagrange reverse trajectory. Using the over-standard sensor as the endpoint, real-time meteorological data such as wind speed, wind direction, temperature, and humidity are input, and the fourth-order Runge-Kutta method is used to solve the reverse trajectory equation. The expression for the reverse trajectory equation is:

[0051]

[0052]

[0053] in, , , For polluting particles at any time Position and time; , , The coordinates of the sensor exceeding the standard; For the variable in the integration time The real-time position coordinates of the contaminating particles; , , These represent the eastward wind speed and volume, northward wind speed component, and vertical wind speed and volume components of the three-dimensional wind speed components, output based on meteorological research and forecasting models. The integration step size is dynamically adjusted within the range of 0.5–5 seconds based on real-time wind speed. Three pollution transmission paths are inverted, and the intersection area is the candidate pollution source area.

[0054] The second stage involves fine-tuning the multi-head graph attention network. A 3D spatial mesh is constructed with the candidate regions as constraints. Node features are input, and node weights are calculated using the multi-head graph attention network. The expression for the attention coefficient is:

[0055]

[0056] in, For the first A grid under attention With neighboring grid Attention coefficient; This is the normalization function; For activation functions; For the first The learnable weight matrix for each attention head; , The feature vector of each grid node contains the concentration value, distance from the assumed pollution source, wind speed component, and building obstruction coefficient. The terrain slope, after being normalized, is then input. This indicates the concatenation of eigenvectors; For grid The output characteristics, after function Mapped to the probability of pollution sources; The number of attention heads is set to 8. For nodes The neighborhood set. Output a probability cloud map from 0 to 1. Regions with a probability ≥ 0.8 are identified as high-suspect pollution source regions. The single inference delay does not exceed 500ms.

[0057] Subsequently, the pollutant component proportion vectors of high-suspect areas were extracted and matched with the historical emission fingerprint database of enterprises in the industrial park using cosine similarity. The cosine similarity expression is:

[0058] in, Fingerprint similarity of pollution source emission indicators; This is the emission fingerprint vector of the pollution source to be measured in real time; For historical emissions fingerprint vectors of companies in the park; It is the dot product of the real-time emission fingerprint vector of the pollution source to be measured and the historical emission fingerprint vector of the enterprises in the park; are the L2 norms of the fingerprint vector to be tested and the historical fingerprint vector, respectively. When A value ≥0.90 indicates a successful match and confirmation as a highly suspected pollution source. The entire source tracing process takes ≤60 seconds, with a positioning error ≤10 meters.

[0059] The hybrid intelligent engine layer outputs diffusion inference results according to the following rules: when the prediction duration is ≤30 minutes, a surrogate model combining a convolutional neural network and a long short-term memory network (CNN-LSTM) is prioritized for rapid inference; when the prediction duration is >30 minutes, a CFD model (three-dimensional convection-diffusion equation) is activated, and the surrogate model results are physically corrected every 10 minutes; when the wind speed is >8 m / s or the wind direction swing is >45°, the mechanistic model is automatically switched. The output results of the surrogate model and the mechanistic model are compared at set time points, and an automatic alarm is triggered when the difference is >25%.

[0060] The autonomous control layer establishes bidirectional communication connections with the hybrid intelligent engine layer, the digital twin construction layer, and the data interconnection and spatiotemporal synchronization layer. This layer takes the source tracing results, diffusion inference results, and the real-time state output from the digital twin construction layer as input, and uses Deep Deterministic Policy Gradient Reinforcement Learning (DDPG) to generate the control strategy. The state space includes sensitive point concentration, equipment operating status, meteorological parameters (wind speed, wind direction, temperature, humidity), diffusion range, and terrain correction coefficient. and building shading coefficient The action space includes spray angle (0~90°), fan frequency (10~50Hz), and valve opening (0~100%).

[0061] The expression for the reward function is:

[0062] in, For comprehensive risk scoring; Real-time concentration at sensitive points; The concentration threshold specified by ambient air quality standards; To control equipment operating costs, the accounting dimensions include equipment energy consumption, chemical consumption, and maintenance losses; The baseline cost based on historical operational data; The rate of decrease in concentration; Indicator function. When The value is 1 when ≥60%, otherwise it is 0; 0.7, 0.3, and 0.2 are the weights of the concentration item, energy consumption item, and treatment efficiency item, respectively. The adjustment method is to increase the weight of the concentration item by 0.05 to 0.10, correspondingly decrease the weight of the energy consumption item, and keep the weight of the treatment efficiency item unchanged. This adjustment rule is determined based on 1,000 digital twin virtual simulation experiments and 5 real leakage events. When multiple pollution sources occur simultaneously, a comprehensive control priority scoring model is used to determine the control sequence. The expression for the comprehensive control priority scoring model is:

[0063] in, The priority of comprehensive management and control of pollution sources is scored, and the higher the score, the higher the priority of management and control of the pollution source. The value represents the toxicity and source strength risk, ranging from 0 to 1, and is determined based on the list of hazardous chemicals and the technical guidelines for environmental impact assessment. This is the distance response value of the sensitive point, which ranges from 0 to 1 and is inversely proportional to the distance to residential areas, schools, and hospitals. The leakage persistence index ranges from 0 to 1 and is determined based on the leakage duration (value 1 for 30 minutes or more) and leakage rate; 0.6, 0.3, and 0.1 are the weights of the three indicators of risk, sensitivity, and urgency, respectively.

[0064] Automatic scheduling and control of equipment is based on priority scores, from highest to lowest. The control resource allocation rule is as follows: Level 1 ( ≥8.5) Call all devices, maximum intensity; Level 2 (6.0≤ <8.5) Utilize 70% of equipment, medium to high governance intensity; Level 3 (3.0≤ <6.0) 40% of devices are used, medium intensity; Level 4 ( <3.0) Scheduled inspections plus intermittent treatment, and 20% of devices are activated. The autonomous control layer automatically encapsulates the strategy into Modbus TCP (Industrial Ethernet Protocol) or OPC UA (Industrial Interoperability Unified Architecture) protocol instructions, and sends them to the executable control devices in the physical space layer through the data interconnection and spatiotemporal synchronization layer, with an end-to-end latency of no more than 10 seconds.

[0065] The autonomous control layer also achieves closed-loop evaluation and optimization based on the concentration decrease rate. (Concentration decrease rate) The expression is:

[0066] in, To control the average concentration within 1 minute prior to the issuance of the command; This is to control the average concentration collected during the period of stable effect (usually 3–10 minutes) after the command is executed. Maintain the current control strategy when ≥60%; when 30% ≤ When the concentration is less than 60%, the weight parameters in the gradient reward function of the deep deterministic strategy are automatically fine-tuned. These weight parameters are the concentration term weight (0.7), energy consumption term weight (0.3), and governance efficiency term weight (0.2) in the reward function. The adjustment method is to increase the concentration term weight by 0.05 to 0.10, correspondingly decrease the energy consumption term weight, and keep the governance efficiency term weight unchanged. This adjustment rule is determined based on 1000 digital twin virtual simulation experiments and post-mortem training of 5 real leakage events. When the concentration is less than 30%, the secondary enhanced control will be automatically activated, including increasing the spray pressure to 1.2 times, increasing the fan frequency to 80% of the rated value, and closing unnecessary valves around the pollution source.

[0067] The digital sand table simulation module establishes a two-way communication connection with the digital twin construction layer and the hybrid intelligent engine layer. The digital twin construction layer outputs real-time 3D scenes, equipment status, terrain parameters, and building parameters to the digital sand table; the hybrid intelligent engine layer outputs source tracing algorithms, diffusion models, and historical case data to the digital sand table. The digital sand table outputs simulated leakage parameters and simulated meteorological conditions to the digital twin construction layer for twin model training and calibration, and outputs simulated strategy execution results to the hybrid intelligent engine layer for optimizing source tracing algorithms and diffusion model parameters. The digital sand table supports custom editing of leakage locations, pollution source strength, meteorological conditions, and terrain; it supports manually setting or calling the autonomous control layer to optimize strategies for emergency strategy rehearsals; it supports multi-person collaborative drills involving command, operation, and monitoring roles; and it supports full-process data recording and automatically generates debriefing reports including timelines, strategy analysis, and effect comparisons.

[0068] Based on the aforementioned digital twin-driven odor source tracing control and simulation system, this invention also provides a digital twin-driven odor source tracing control and simulation method, such as... Figure 2 As shown, it includes the following steps: S1. Raw Data Acquisition and Upload: Raw data on odor concentration, meteorological parameters, and equipment status are collected through the physical space layer and uploaded to the data interconnection and spatiotemporal synchronization layer. The deployment rules of the multimodal sensing network in the physical space layer are as follows: sensors are deployed 30m to 80m downwind of the pollution source, at the factory boundary, and at corridor turning points; sensors are deployed 50m to 100m upwind of sensitive points (residential areas, schools, hospitals); key control areas (tank areas, reactor areas) are deployed in 50m*50m grids, and general areas are deployed in 100m*100m grids.

[0069] S2, Data Cleaning and Spatiotemporal Alignment The data interconnection and spatiotemporal synchronization layer cleanses and spatiotemporally aligns the raw data, then outputs standardized data to the digital twin construction layer; among other things, by adopting... Outliers are removed using criteria, a moving average window is used for smoothing filtering, and millisecond-level Coordinated Universal Time timestamps are uniformly aligned with the WGS84 coordinate system.

[0070] S3, Real-time Concentration Field and Meteorological Field Generation The digital twin construction layer generates real-time concentration and meteorological fields based on standardized data and pre-stored 3D models, and outputs them to the hybrid intelligent engine layer. In this step, a 1:1 3D model is constructed based on GIS (Geographic Information System) data, BIM (Building Information Modeling) data and laser point cloud data. An improved Gaussian plume model, a CFD (3D convection-diffusion equation) model and a convolutional neural network and long short-term memory network (CNN-LSTM) surrogate model are integrated to output the concentration field, meteorological field and correction coefficients.

[0071] S4, Two-Stage Source Tracing and Hybrid Driven Deduction First, candidate regions are obtained by inverting the Lagrange inverse trajectory using the over-standard sensor as the endpoint. Then, a multi-head image attention network (number of attention heads) is used. =8) Calculate the pollution source probability for the spatial grid within the candidate region and output a probability cloud map. Regions with a probability ≥ 0.8 are identified as high-suspect pollution source regions. Subsequently, the emission fingerprint matching verification is performed using the cosine similarity calculation formula. A value ≥0.90 indicates a highly suspected pollution source. Simultaneously, based on the prediction duration, either a proxy model or a mechanistic model is selected for diffusion simulation: when the prediction duration is ≤30 minutes, the proxy model is prioritized; when the prediction duration is >30 minutes, the mechanistic model is periodically corrected; and in complex scenarios (wind speed >8 m / s or wind direction swing exceeding 45°), the system automatically switches to the mechanistic model.

[0072] S5. Autonomous control command issuance The autonomous control layer acquires the source tracing results, simulation results, and the real-time status of the twin. It generates control policies through Deep Deterministic Policy Gradient Reinforcement Learning (DDPG) and issues commands to the executable control devices in the physical space layer according to priority via the data interconnection and spatiotemporal synchronization layer. In this step, priority is determined based on the integrated management and control priority scoring model. Confirmed. The instruction is encapsulated in Modbus TCP or OPCUA protocol format, with an end-to-end delay of ≤10s.

[0073] S6. Evaluation of Governance Effectiveness and Adjustment of Strategies After the executable control device executes the command, the autonomous control layer collects the concentration data after treatment and calculates the concentration reduction rate. And assess the effectiveness of the treatment based on the rate of decline. When Maintain the current control strategy when ≥60%, and when 30% ≤ When <60%, the strategy is automatically adjusted. When the spray level is less than 30%, the secondary enhanced control will be automatically activated (including increasing the spray pressure, increasing the fan frequency, and closing unnecessary peripheral valves).

[0074] S7, Virtual Scene Construction and Data Feedback The digital sandbox simulation module receives real-time data from the digital twin construction layer and the hybrid intelligent engine layer to build a virtual scene, and feeds back the simulated leakage parameters and strategy execution results to the digital twin construction layer and the hybrid intelligent engine layer for model training and parameter optimization.

[0075] Example Using a chemical industrial park (approximately 20 square kilometers) as the implementation target, the digital twin-driven odor source tracing control and simulation exercise system and method of this invention were fully verified.

[0076] First, a multimodal sensing network was deployed within the park according to the following deployment rules: sensors were deployed 30m–80m downwind of pollution sources, at the factory boundary, and at corridor turning points; sensors were deployed 50m–100m upwind of sensitive points (residential areas, schools, hospitals); key control areas (tank areas, reactor areas) were deployed using a 50m*50m grid, and general areas were deployed using a 100m*100m grid. Based on these rules, a total of 80 gas sensors were deployed (32 in high-risk grids such as key pollution sources and tank areas, and 48 in grids around sensitive points and in general monitoring areas), 12 mini weather stations, 8 video surveillance systems, and 4 inspection drones were deployed. Simultaneously, executable control equipment was installed at key locations, including 16 high-pressure spray devices, 12 variable frequency fans, and 24 smart valves.

[0077] Through a data interconnection and spatiotemporal synchronization layer, a message queue telemetry transmission protocol is used to achieve real-time data synchronization. Concentration data and meteorological data are uploaded at a frequency of once per second, while equipment status data is uploaded at a frequency of once every 2 seconds. This layer cleans the received raw data and uses... Outliers are removed using the criteria, and smoothing is performed using a moving average filter. Then, all data is unified to millisecond-level Coordinated Universal Time timestamps and the WGS84 coordinate system to achieve spatiotemporal alignment.

[0078] In the digital twin construction layer, based on the park's GIS (Geographic Information System) data, BIM (Building Information Modeling) data, and laser point cloud data (point cloud density not less than 100 points / m²), 2 A 1:1 3D scene is constructed, with model coordinate errors controlled within 0.5m. This layer integrates an improved Gaussian plume model, a CFD model (3D convection-diffusion equation), and a convolutional neural network and long short-term memory network (CNN-LSTM) surrogate model, outputting real-time concentration fields, meteorological fields, and terrain and building occlusion correction coefficients. This achieves real-time synchronous mapping between the physical park and the digital twin, with a typical synchronization delay ≤1s.

[0079] At 14:23:05 on a certain workday, gas sensors No. 36, 38, and 41 near the sensitive point on the east side of the park detected excessive concentrations of volatile organic compounds (VOCs) for three consecutive sampling cycles (1 second per cycle), with a peak concentration of approximately 150 μg / m³. 3 The system automatically triggers an emergency response. Within 1 second, the data interconnection and spatiotemporal synchronization layer synchronizes the excessive data to the digital twin construction layer. The corresponding point in the digital twin in the 3D scene flashes red as a warning, and an automatic warning circle with a range of 50m is generated. At the same time, the nearest drone is automatically dispatched to take off and head to the suspicious area.

[0080] The hybrid intelligent engine layer then initiates a two-stage source tracing mechanism. In the first stage, using sensors 36, 38, and 41 as endpoints, the fourth-order Runge-Kutta method is employed to solve the Lagrange inverse trajectory equation. The integration step size is dynamically adjusted based on real-time wind speed (approximately 2.5 m / s to 3.5 m / s), resulting in three pollution transmission paths. The intersection of these paths points to the tank area of ​​a chemical company within the industrial park, completing the coarse localization. In the second stage, using this candidate area as a constraint, a three-dimensional spatial mesh is constructed. Input node features (including concentration values, distance to the assumed pollution source, wind speed and direction, building obstruction coefficient, etc.) are used. A multi-head graph attention network (number of attention heads) is then used to... =8) Calculate the pollution source probability for each grid and output a probability cloud map, where the probability value for a specific tank area in the tank farm reaches 0.89. Subsequently, extract the pollutant component proportion vector for this area and perform cosine similarity matching with the historical emission fingerprint database of enterprises in the industrial park to calculate... =96%, confirming the company's storage tank as the source of the leak. The entire tracing process took approximately 55 seconds, with a location error of less than 10 meters.

[0081] Simultaneously, the hybrid intelligent engine layer initiates a diffusion simulation. Since the prediction timeframe is 15 minutes (considered a short-term prediction), the system prioritizes using a convolutional neural network and a long short-term memory network (CNN-LSTM) surrogate model for rapid simulation. The simulation results show that after 15 minutes, the pollution cloud will spread to the downwind boundary of the residential area, with a peak concentration of approximately 180 μg / m³. 3 Subsequent verification based on actual monitoring data showed that the average error of the prediction was approximately 6.2%, and the entire simulation process was completed within milliseconds.

[0082] The autonomous control layer then automatically obtains the following inputs from the digital twin construction layer and the hybrid intelligent engine layer: the leak source is located in a tank area of ​​a certain enterprise, the source strength is approximately 50 μg / s, the current weather conditions are a southeast wind with a speed of 3.2 m / s, and the current concentration in the residential area is 120 μg / m³. 3 The peak concentration in the residential area is predicted to reach 180 μg / m³ after 15 minutes. 3The system monitors real-time equipment status, terrain correction coefficients, and building obstruction coefficients. A Deep Deterministic Policy Gradient Reinforcement Learning (DDPG) model is used to generate control strategies. The concentration term in the model's reward function has a weight of 0.7, reflecting the goal of prioritizing air quality at sensitive points. The generated strategy is to adjust the angle of the upwind high-pressure spray device to 45°, increase the frequency of the downwind variable frequency fan to 45Hz, and close three smart valves around the leak source. The autonomous control layer automatically encapsulates these strategies into Modbus TCP protocol commands, which are then transmitted to the corresponding execution devices in the physical space layer via data interconnection and spatiotemporal synchronization layers, with an end-to-end latency of approximately 7 seconds. The high-pressure spray device, variable frequency fan, and smart valves respond within 1–5 seconds of receiving the commands and begin executing the control actions.

[0083] Approximately 8 minutes after the control command was executed (the stabilization period determined based on the pollution plume's diffusion rate), the autonomous control layer collected concentration data from the treated residential area and calculated the concentration reduction rate. The average concentration before control was 120 μg / m³. 3 The controlled average concentration was 62 μg / m³. 3 The concentration decrease rate was calculated. The concentration was approximately 48.3%. Since this value falls between 30% and 60%, the autonomous control layer automatically fine-tunes the weight parameters in the reinforcement learning reward function and generates a new optimization strategy. After implementing the optimization strategy, the concentration in the residential area further decreased to 52 μg / m³. 3 Concentration decrease rate The efficiency rate increased to 56.7%, at which point the system determined that the governance effect met the requirements and maintained the current control strategy. The entire process required no manual intervention, achieving a fully autonomous closed loop from perception, decision-making, execution to evaluation and optimization.

[0084] To verify the long-term operational performance of the system, a continuous seven-month operational verification was conducted in the aforementioned chemical industrial park (March 2024 to September 2024), during which 12 real-world leakage events and 100 virtual simulation scenarios were performed. The operational data is summarized in Table 1 below: Table 1. Verification of the overall performance and implementation effect of this system

[0085] The above verification results show that the system in this invention is significantly better than the traditional system in key performance indicators such as tracing time, positioning error, prediction error, control command delay, closed-loop processing time, evaluation accuracy, inference delay and priority matching accuracy. The system's operational stability reaches over 99.6%, demonstrating good engineering feasibility.

[0086] In summary, the digital twin-driven odor source tracing control and simulation exercise system and method provided by this invention, through the collaborative architecture of the physical space layer, data interconnection and spatiotemporal synchronization layer, digital twin construction layer, hybrid intelligent engine layer, autonomous control layer and digital sandbox simulation exercise module, and the complete method flow based on this architecture, achieves the organic integration of high-precision source tracing, autonomous closed-loop control without human intervention and virtual-real feedback simulation exercise in the field of open space odor pollution control.

Claims

1. A digital twin-driven odor source tracing control and simulation system, characterized in that: It includes a physical space layer, a data interconnection and spatiotemporal synchronization layer, a digital twin construction layer, a hybrid intelligent engine layer, an autonomous control layer, and a digital sandbox simulation exercise module; The physical space layer includes a multimodal sensing network and executable control devices for collecting raw data including concentration, meteorology, and equipment status. The data interconnection and spatiotemporal synchronization layer is bidirectionally connected to the physical space layer, and is used to receive the raw data, clean and spatiotemporally align the raw data to obtain standardized data, transmit the standardized data outward, and receive control commands and send them to the physical space layer. The digital twin construction layer is unidirectionally connected to the data interconnection and spatiotemporal synchronization layer, and is used to construct a three-dimensional scene based on geographic information, building information model and laser point cloud, and integrate the mechanism model and data-driven proxy model to output real-time concentration field, meteorological field and correction coefficient. The hybrid intelligent engine layer is unidirectionally connected to the digital twin construction layer. It is used to output the source tracing results by adopting a two-stage mechanism of first coarse localization of the Lagrange reverse trajectory and then correction by the multi-head graph attention network. It also adaptively calls the proxy model or mechanism model according to the prediction duration and meteorological conditions, and outputs the diffusion inference results by the rules of the mechanism model that are periodically corrected. The autonomous control layer is bidirectionally connected to the hybrid intelligent engine layer, the digital twin construction layer, and the data interconnection and spatiotemporal synchronization layer. It is used to generate control strategies through reinforcement learning by taking the source tracing results, the diffusion inference results, and the real-time status output by the digital twin construction layer as inputs, and to issue control commands to the data interconnection and spatiotemporal synchronization layer according to priority. It also performs evaluation and adaptive optimization of the strategy based on the governance effect. The digital sand table simulation exercise module is bidirectionally connected to the digital twin construction layer and the hybrid intelligent engine layer, and is used to construct virtual leakage scenarios, rehearse emergency strategies, and conduct bidirectional interaction of virtual and real data with the digital twin construction layer and the hybrid intelligent engine layer.

2. The digital twin-driven odor source tracing control and simulation system according to claim 1, characterized in that: The improved Gaussian plume model integrated in the digital twin construction layer introduces a terrain correction coefficient. With building shading correction factor Its expression is: in, The pollution source is strong; Average wind speed; For lateral diffusion parameters; For vertical diffusion parameters; High efficiency of source; This is a terrain correction factor, determined by ground roughness and slope; The building shading correction factor is determined by the building's average height, surface density, and the distance from the calculation point to the building's centerline.

3. The digital twin-driven odor source tracing control and simulation system according to claim 1, characterized in that: In the hybrid intelligent engine layer, the Lagrange reverse trajectory coarse localization uses the fourth-order Runge-Kutta method to solve the reverse trajectory equation, and the integration step size is dynamically adjusted according to the real-time wind speed. The overall source tracing time is ≤60s. The multi-head graph attention network correction uses 8 attention heads, and areas with an output probability ≥0.8 are identified as high-suspect pollution source areas.

4. The digital twin-driven odor source tracing control and simulation system according to claim 1, characterized in that: The autonomous control layer employs deep deterministic policy gradient reinforcement learning for decision-making, and its reward function is: in, For comprehensive risk scoring; Real-time concentration at sensitive points; The concentration threshold specified by ambient air quality standards; To control equipment operating costs, the accounting dimensions include equipment energy consumption, reagent consumption, and maintenance losses; The baseline cost based on historical operational data; The rate of decrease in concentration; Indicator functions.

5. The digital twin-driven odor source tracing control and simulation exercise system according to claim 1, characterized in that: The autonomous control layer uses the concentration decrease rate Achieve closed-loop evaluation and optimization, concentration reduction rate The expression is: in To control the average concentration before the instruction is issued; To control the average concentration after the command is executed; For the rate of decrease in concentration, when When the level is less than 30%, the secondary enhanced control will be automatically activated.

6. The digital twin-driven odor source tracing control and simulation system according to claim 1, characterized in that: The autonomous control layer also includes a comprehensive management and control priority scoring module, which is used to calculate the management and control priority score of each pollution source when multiple pollution sources occur simultaneously. Its expression is: in, Priority scoring for comprehensive pollution source management; Risk values ​​for toxicity and source strength; The distance response value is the distance from the sensitive point. The leakage persistence index is represented by 0.6, 0.3, and 0.1, which are the weights of the three indicators: risk, sensitivity, and urgency, respectively. The autonomous control layer scores based on the control priority. The system automatically allocates control resources of different proportions and intensities based on the level of the system.

7. The digital twin-driven odor source tracing control and simulation system according to claim 1, characterized in that: The deployment rules for the multimodal sensing network in the physical space layer are as follows: sensors are deployed 30m to 80m downwind of the pollution source, at the factory boundary, and at the inflection point of the corridor; sensors are deployed 50m to 100m upwind of sensitive points; key control areas are deployed in 50m*50m grids, and general areas are deployed in 100m*100m grids.

8. The digital twin-driven odor source tracing control and simulation exercise system according to claim 1, characterized in that: The bidirectional data communication rules between the digital sand table simulation exercise module, the digital twin construction layer, and the hybrid intelligent engine layer are as follows: the digital twin construction layer outputs real-time 3D scenes and device status to the digital sand table, and the hybrid intelligent engine layer outputs source tracing algorithms and diffusion models to the digital sand table; the digital sand table outputs simulated leakage parameters to the digital twin construction layer for model calibration, and outputs simulated strategy execution results to the hybrid intelligent engine layer for algorithm optimization.

9. A digital twin-driven method for odor source tracing control and simulation, characterized in that, The odor source tracing control and simulation system driven by digital twins according to any one of claims 1 to 8 includes the following steps: S1. Raw Data Acquisition and Upload Raw data on odor concentration, meteorological parameters, and equipment status are collected through the physical space layer and uploaded to the data interconnection and spatiotemporal synchronization layer. S2, Data Cleaning and Spatiotemporal Alignment The data interconnection and spatiotemporal synchronization layer cleans and spatiotemporally aligns the original data, and then outputs standardized data to the digital twin construction layer. S3, Real-time Concentration Field and Meteorological Field Generation The digital twin construction layer generates real-time concentration fields and meteorological fields based on the standardized data and pre-stored 3D models, and outputs them to the hybrid intelligent engine layer. S4, Two-Stage Source Tracing and Hybrid-Driven Deduction The hybrid intelligent engine layer initiates a two-stage source tracing: first, it uses the excessive sensor as the endpoint to invert the Lagrange reverse trajectory to obtain candidate areas; then, it uses a multi-head graph attention network to calculate the pollution source probability of the spatial grid within the candidate areas and outputs a probability cloud map; at the same time, it selects to call a proxy model or a mechanism model to perform diffusion inference based on the prediction duration. S5. Autonomous control command issuance The autonomous control layer acquires the tracing results, inference results, and real-time status of the twin, generates control strategies through reinforcement learning, and issues instructions to the executable control devices of the physical space layer according to priority via the data interconnection and spatiotemporal synchronization layer. S6. Evaluation of Governance Effectiveness and Adjustment of Strategies After the executable control device executes the command, the autonomous control layer collects the concentration data after treatment, calculates the concentration reduction rate, and evaluates the treatment effect based on the concentration reduction rate. If the requirements are not met, the strategy is automatically adjusted or a secondary enhanced control is initiated. S7, Virtual Scene Construction and Data Feedback The digital sandbox simulation module receives real-time data from the digital twin construction layer and the hybrid intelligent engine layer to construct a virtual scene, and feeds back the simulated leakage parameters and strategy execution results to the digital twin construction layer and the hybrid intelligent engine layer for model training and parameter optimization.

Images