Ship exhaust gas diffusion field reconstruction method and system based on physical information fusion driving

Abstract

The application discloses a ship exhaust diffusion field reconstruction method and system based on physical information fusion driving, which comprises the following steps: acquiring multi-source input data of a target inland waterway region; constructing a physical residual based on a three-dimensional convection-diffusion-reaction equation, embedding the physical residual into a deep neural network training framework; explicitly introducing physical constraints of entity obstacle boundaries and underlying surface settlement boundaries in the neural network training process, constructing a block mixed boundary constraint mechanism; fitting losses based on sparse observation data, physical residual loss, mixed boundary constraint loss, introducing a neural tangent kernel adaptive weighting strategy, constructing a multi-task joint loss function, and training the neural network model; and the neural network model reconstructs a ship exhaust diffusion concentration field according to real-time input multi-source input data. The application can realize stable and rapid reconstruction of a ship exhaust diffusion field in a complex limited water area.

Description

Technical Field

[0001] This invention relates to the field of atmospheric pollutant diffusion monitoring and reconstruction technology, specifically to a method and system for reconstructing ship exhaust diffusion fields based on physical information fusion. Background Technology

[0002] Monitoring the dispersion of ship exhaust emissions is a common challenge in global environmental governance. The scientific prerequisite for accurate ship emission monitoring lies in the ability to quickly and accurately characterize the spatiotemporal evolution of atmospheric pollutants in the real environment. Unlike open seas, inland waterways are constrained by numerous solid structures such as bridge piers and shoreline buildings. These solid obstacles mechanically obstruct and bypass exhaust plumes, causing the micro-meteorological field of near-shore waterways to exhibit significant non-steady-state characteristics, resulting in a highly nonlinear and uncertain ship exhaust dispersion process. From the perspective of diffusion theory, the widely used Gaussian diffusion model and its derivatives are based on the idealized assumptions of a homogeneous meteorological field, isotropic turbulence, and infinite boundaries. However, in complex boundary scenarios such as inland waterways, these assumptions are difficult to uphold.

[0003] In summary, the diffusion of ship exhaust gas in inland waterways is essentially a multi-scale physical process simultaneously dominated by strong topological boundary constraints and non-steady-state micro-meteorological field disturbances, and its dynamic characteristics differ significantly from those in ports or open waters. This difference leads to: (1) at the mechanistic level, traditional diffusion models based on the uniformity assumption are difficult to characterize the nonlinear modulation mechanism of pollutant diffusion by boundary-induced non-equilibrium turbulence; (2) at the methodological level, the computational efficiency of traditional numerical diffusion models cannot meet the requirements for rapid inversion of diffusion concentration fields under multiple scenarios and uncertain conditions required for monitoring.

[0004] Existing methods have the following main technical shortcomings in solving the above problems: (1) Mechanism-driven methods (Gaussian model, AERMOD, CALPUFF, and WRF-Chem, etc.): Their core assumptions are based on the idealized premise of a uniform meteorological field, isotropic turbulence, and infinite boundaries. However, the mechanical obstruction and flow-around effects of physical structures such as bridge piers and shoreline buildings (inducing canyon winds and topographic backflow zones) on exhaust plumes in inland waterways fundamentally violate the above ideal assumptions, resulting in prediction errors of several times or more in complex scenarios such as bridge areas using traditional Gaussian models. In addition, although high-precision computational fluid dynamics / large eddy simulation can characterize complex boundaries, simulation of a single scenario takes several hours to several days and cannot be run in real time on shore-based monitoring systems, thus its engineering practicality is extremely limited.

[0005] (2) Pure data-driven methods (support vector machine, artificial neural network, etc.): These methods are highly dependent on the distribution of training data and have insufficient generalization ability when encountering meteorological conditions outside the training set. More importantly, the lack of physical constraints makes the model prone to non-physical predictions near physical obstacles such as bridge piers and buildings (i.e., the concentration distribution does not conform to the physical conservation law), which seriously affects the credibility of the results and cannot be used to guide actual regulatory decisions.

[0006] (3) In recent years, physics-data fusion methods have gradually become a research hotspot. Among them, Physics-Informed Neural Networks (PINNs) provide a new approach to solving complex atmospheric diffusion prediction problems by introducing physical constraints during the learning process. PINNs improve computational efficiency while maintaining physical consistency and have shown potential in fluid dynamics and reaction-diffusion systems. However, their stability, boundary handling ability, and optimization convergence characteristics in complex boundary and multi-scale unsteady atmospheric diffusion problems still need further investigation.

[0007] In summary, there is an urgent need for a method for rapid reconstruction of ship exhaust gas diffusion fields that can run in real time on a shore-based edge computing device equipped with a processor, combining physical consistency of physical boundaries with computational efficiency, and can form an engineering closed loop with shore-based IoT monitoring systems and ship regulatory decision-making systems. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for reconstructing ship exhaust diffusion fields based on physical information fusion, which can achieve stable and rapid reconstruction of ship exhaust diffusion fields in complex and confined waters, and is especially suitable for inland waterway environments with complex boundary constraints such as bridges and shoreline structures.

[0009] The technical solution adopted in this invention is: A method for reconstructing the ship exhaust gas diffusion field based on physical information fusion includes the following steps: Multi-source input data for the target inland waterway area is acquired through a shore-based Internet of Things (IoT) system. The multi-source input data includes: three-dimensional meteorological field data, ship emission source data, and measured pollutant concentration data detected by shore-based sensors. Physical residuals are constructed based on three-dimensional convection-diffusion-reaction equations, and the physical residuals are embedded into a deep neural network training framework using automatic differentiation techniques. During neural network training, physical constraints of physical obstacle boundaries and underlying surface settlement boundaries are explicitly introduced to construct a block-based hybrid boundary constraint mechanism. Based on sparse observation data fitting loss, physical residual loss, and hybrid boundary constraint loss, an adaptive weighting strategy of neural tangent kernel is introduced to construct a multi-task joint loss function and train the neural network model. The trained physical information neural network performs forward inference based on real-time input three-dimensional meteorological field data and ship emission source data, quickly reconstructs the ship exhaust gas diffusion concentration field, and generates regulatory instructions based on the concentration field.

[0010] Preferably, the physical residual constructed based on the three-dimensional convection-diffusion-reaction equation is:

[0011] in, For dynamic ship source terms reconstructed based on AIS, X is the ship position vector and t is time; The pollutant concentration field predicted by the neural network. This refers to the gradient of the pollutant concentration field predicted by the neural network; This is a three-dimensional wind field vector; The turbulent diffusion coefficient characterizes the ability of atmospheric turbulence to mix pollutants; The comprehensive settlement coefficient taking into account the effects of rain / fog.

[0012] Preferably, the physical constraints of the solid obstacle boundary are: applying a flux-free constraint to the surface of the solid obstacle in the inland waterway, and the physical constraints of the underlying surface settlement boundary are: applying a settlement constraint to the water surface and / or land surface. The physical obstacles include bridge piers and / or shoreline structures. Neumann-type flux-free constraints are applied to the surfaces of these physical obstacles in inland waterways. ; used to represent the flow around and obstruction of pollutants, avoiding non-physical diffusion results that penetrate buildings from the model, n obs Let be the unit outward normal vector of the boundary surface of the solid obstacle; Apply Robin-type mixed boundary conditions to both water and land surfaces. ,in The dry / wet deposition velocity parameter is used to characterize the flux of pollutants adsorbed and removed by water bodies or land surfaces, effectively alleviating the problem of systematic overestimation of near-surface concentration; D is the turbulent diffusion coefficient, n dep Let be the unit outward normal vector of the settlement boundary surface of the underlying surface. The pollutant concentration field predicted by the neural network. This refers to the gradient of the pollutant concentration field predicted by the neural network.

[0013] Preferably, the multi-task joint loss function is:

[0014] in, R represents the fitting loss for sparse observation data; the data comes from fixed-point sampling of a high-fidelity simulated field or actual measured values; PDE For physical residuals; These are the weighting coefficients for the fitting term of the sparse observation data; n represents the weighting coefficient of the physical residual term; obs n is the unit outward normal vector of the boundary surface of the solid obstacle; dep λ is the unit outward normal vector of the underlying surface at the settlement boundary; b1 This refers to the adaptive weighting coefficient of the entity obstacle boundary; λ b2 This refers to the adaptive weighting coefficient of the underlying surface settlement boundary; D The turbulent diffusion coefficient; The pollutant concentration field predicted by the neural network; This refers to the gradient of the pollutant concentration field predicted by the neural network; These are the dry / wet settling velocity parameters; Indicates the boundary of a physical obstacle; This indicates the settlement boundary of the underlying surface.

[0015] Preferably, during the training iteration process, the gradient spectrum feature value corresponding to each loss component is calculated in real time based on the neural tangent, and the weight coefficient λ is dynamically adjusted according to the gradient spectrum feature value. b1 and / or λ b2 .

[0016] Preferably, the gradient spectrum feature value corresponding to each loss component is calculated in real time based on the neural tangent, and the weight coefficient is dynamically adjusted according to the gradient spectrum feature value. Each loss component refers to the data fitting loss, physical residual loss and mixed boundary constraint loss mentioned above. The specific adjustment steps for dynamically adjusting the weight coefficients based on the gradient spectrum feature values ​​are as follows: when the gradient spectrum feature value of the mixed boundary constraint loss component is detected to be lower than a preset threshold, the corresponding mixed boundary constraint weight coefficient is adaptively increased. Specifically: Let the gradient spectrum eigenvalue of the boundary constraint term in the t-th training iteration be... The preset threshold is ,when When the boundary constraint weight coefficients are updated, they are updated as follows:

[0017] Where η is the weight update step size, ε is the stability term, λmax is the upper limit of the boundary constraint weights, and λbound is the boundary constraint weight coefficient. This method allows the optimization priority of the boundary constraint terms during training to adaptively adjust with changes in their gradient spectrum eigenvalues, thereby improving the model's satisfaction of physical constraints such as obstacle boundaries and depositional boundaries.

[0018] Preferably, the acquisition of the three-dimensional meteorological field data includes the following steps: using a multi-scale nested numerical simulation method to nest the regional-scale background wind field with the local micro-scale flow field, wherein the immersion boundary method is used to represent the non-convex geometric boundary formed by bridge piers and / or shoreline building complexes. The multi-scale nested numerical simulation method includes a one-way coupling strategy of mesoscale meteorological models and computational fluid dynamics models to obtain background wind fields and thermal structures at kilometer-level resolution, which serve as boundary driving conditions for microscale simulations. At the waterway scale, a partial analytical simulation method for turbulence with a resolution of 100 meters was used to obtain the dominant flow pattern and turbulence statistical characteristics of the near-surface layer of inland waterways. Locally encrypted nested domains are set up in bridge areas and densely built areas. For non-convex geometric boundaries formed by building groups and bridge structures, the immersion boundary method is adopted. The volume force source term is introduced into the momentum equation to replace the body-fit solution of the solid structure. The specific process for acquiring ship emission source data is as follows: ship trajectory and operating condition data are collected through the Automatic Identification System (AIS), the instantaneous emission intensity of the ship is inverted, and the Lagrange-Euler hybrid description method is adopted. A Gaussian kernel function is introduced to map discrete moving point sources into a continuous Euler grid three-dimensional dynamic emission source term field.

[0019] Preferably, the neural network model is any one of a fully connected deep neural network, a Fourier feature network, or a deep operator network. The input of the neural network model includes ship emission source data and meteorological environmental data. The ship emission source data includes the spatiotemporal coordinates of the ship emission source and the ship source term. The meteorological environmental data includes wind speed, wind direction, atmospheric stability, rainfall intensity, and / or turbulence diffusion parameters extracted or interpolated from three-dimensional meteorological field data. The neural network model outputs pollutant concentration, and the reconstructed ship exhaust gas diffusion concentration field is obtained by calculating each predicted point in the three-dimensional spatial grid of the measured section.

[0020] Preferably, the ship exhaust gas diffusion field reconstruction method based on physical information fusion further includes the following steps: transmitting the reconstructed concentration field results to a monitoring terminal for visualization.

[0021] A rapid reconstruction system for ship exhaust gas diffusion field based on physical information fusion includes: a shore-based Internet of Things (IoT) system, a monitoring terminal, a processor, and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the ship exhaust gas diffusion field reconstruction method based on physical information fusion is implemented. The shore-based IoT system is connected to the processor and is used to collect multi-source input data and transmit it to the processor. The monitoring terminal is connected to the processor and is used to display the output results of the processor implementing the ship exhaust gas diffusion field reconstruction method based on physical information fusion.

[0022] The beneficial effects of this invention are: This invention enables high-precision and high-efficiency reconstruction of the dispersion concentration field of ship exhaust emissions under complex micrometeorological conditions. It can be widely applied to the precise monitoring of air pollution from inland waterway vessels and the management of green shipping. Through a three-layer mechanism of "deep embedding of physical mechanisms, block-mixed boundary constraints, and adaptive control of neural tangent kernel gradients", it solves the core technical problems of existing methods such as poor physical consistency, unstable training, and insufficient real-time performance in complex physical boundary environments in inland waterways. It achieves stable and rapid reconstruction of the ship exhaust dispersion field in complex and restricted waterways, and is especially suitable for inland waterway environments with complex boundary constraints such as bridges and shoreline structures. Attached Figure Description

[0023] Figure 1 This is a flowchart of the ship exhaust gas diffusion field reconstruction method based on physical information fusion in an embodiment of the present invention; Figure 2 This is an architecture diagram of a rapid reconfiguration system for ship exhaust gas diffusion field driven by physical information fusion in an embodiment of the present invention.

[0024] Figure 3 This is a three-dimensional view of the boundary constraints of the inland river entity in an embodiment of the present invention.

[0025] Figure 4 This is a side view of the boundary constraints of the inland river entity in an embodiment of the present invention.

[0026] Figure 5 This is a top view of the boundary constraints of the inland river entity in an embodiment of the present invention.

[0027] In the picture: 1-bridge pier; 2-shore building complex; 3-ship. Detailed Implementation

[0028] 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 merely illustrative and not intended to limit the invention.

[0029] In the description of this invention, it should be understood that if terms such as "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise" are used to indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, they are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0030] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection. They can refer to a mechanical connection or an electrical connection. They can refer to a direct connection or an indirect connection through an intermediate medium, and they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0031] Example 1 A method for reconstructing the ship exhaust gas diffusion field based on physical information fusion includes the following steps: Step S1: Acquire multi-source input data of the target inland waterway area through the shore-based Internet of Things system; the multi-source input data includes at least: three-dimensional meteorological field data, ship emission source data obtained by inversion from Automatic Identification System (AIS) data, and measured pollutant concentration data detected by shore-based sensors; Step S2: Construct physical residuals based on the three-dimensional convection-diffusion-reaction equations, and embed the physical residuals into a deep neural network training framework using automatic differentiation technology; Step S3: Explicitly introduce physical obstacle boundaries during neural network training. Physical constraints and underlying surface settlement boundaries Physical constraints are used to construct a block-based hybrid boundary constraint mechanism; Step S4: Based on the sparse observation data fitting loss, physical residual loss, and mixed boundary constraint loss, an adaptive weighting strategy of neural tangent kernel is introduced to construct a multi-task joint loss function and train the neural network model. Step S5: Deploy the trained neural network model on an edge computing device, and perform forward inference based on the real-time input three-dimensional meteorological field data and ship emission source data to reconstruct the ship exhaust gas diffusion concentration field.

[0032] Furthermore, the shore-based IoT system includes meteorological sensors deployed at shore-based monitoring stations, Automatic Identification System (AIS) for ships, and multi-point sensor arrays arranged at shore-based fixed monitoring IoT nodes.

[0033] Furthermore, the physical residual constructed in step S2 based on the three-dimensional convection-diffusion-reaction equation is:

[0034] in, For dynamic ship source terms reconstructed based on AIS, X is the ship position vector and t is time; The pollutant concentration field predicted by the neural network. This refers to the gradient of the pollutant concentration field predicted by the neural network; This is a three-dimensional wind field vector; The turbulent diffusion coefficient characterizes the ability of atmospheric turbulence to mix pollutants; The comprehensive settlement coefficient taking into account the effects of rain / fog.

[0035] Furthermore, For advection transport, For turbulent diffusion, This is due to wet deposition attenuation.

[0036] Furthermore, in step S3, the physical constraints of the solid obstacle boundary are: applying a flux-free constraint to the surface of the solid obstacle in the inland waterway, and the physical constraints of the underlying surface settlement boundary are: applying a settlement constraint to the water surface and / or land surface. The physical obstacles include bridge piers and / or shoreline structures. Neumann-type flux-free constraints are applied to the surfaces of these physical obstacles in inland waterways. ; used to represent the flow around and obstruction of pollutants, avoiding non-physical diffusion results that penetrate buildings from the model; n obs This refers to the unit outward normal vector of the boundary surface of a solid obstacle, which can also be understood as "the direction vector perpendicular to the surface of the bridge pier or the outer surface of the shoreline structure". This means that the gradient of the predicted pollutant concentration field in the boundary normal direction is zero, that is, the pollutants do not generate normal diffusion flux on the surface of bridge piers and / or shoreline buildings, thus constraining the model prediction results to meet the physical boundary condition that the physical obstacle is impenetrable. Apply Robin-type mixed boundary conditions to both water and land surfaces. ,in The dry / wet deposition velocity parameter is used to characterize the flux of pollutants adsorbed and removed by water bodies or land surfaces, effectively alleviating the problem of systematic overestimation of near-surface concentration; D is the turbulent diffusion coefficient, n dep Let be the unit outward normal vector of the settlement boundary surface of the underlying surface. The pollutant concentration field predicted by the neural network. This refers to the gradient of the pollutant concentration field predicted by the neural network.

[0037] This paper proposes to achieve precise embedding of block-based hybrid boundary physical constraints in the training framework of physical information neural networks, thereby fundamentally solving the problem of non-physical penetration prediction caused by a single boundary assumption in complex topological environments in existing physical information neural network methods.

[0038] Furthermore, the multi-task joint loss function mentioned in step S4 is:

[0039] in, R represents the fitting loss for sparse observation data, which refers to three-dimensional meteorological field data and ship emission source data; PDE For physical residuals; The weighting coefficients of the fitting term for sparse observation data that varies over time; The weighting coefficients for the time-varying physical residuals; n obs n is the unit outward normal vector of the boundary surface of the solid obstacle; dep λ is the unit outward normal vector of the underlying surface at the settlement boundary; b1 (t) refers to the time-varying adaptive weighting coefficient of the physical obstacle boundary; λ b2 (t) refers to the adaptive weighting coefficient of the underlying surface settlement boundary that varies with time; D The turbulent diffusion coefficient; The pollutant concentration field predicted by the neural network; This refers to the gradient of the pollutant concentration field predicted by the neural network; These are the dry / wet settling velocity parameters; Indicates the boundary of a physical obstacle; This indicates the settlement boundary of the underlying surface.

[0040] Furthermore, step S4 also includes: during the training iteration process, calculating the gradient spectrum feature values ​​corresponding to each loss component in real time based on the neural tangent kernel (NTK), and dynamically adjusting the weight coefficient λ according to the gradient spectrum feature values. b1 and / or λ b2 .

[0041] Furthermore, the gradient contribution of each loss component is evaluated in real time using the eigenvalues ​​of the neural tangent kernel, and the weight coefficients are dynamically adjusted based on the proportion of the eigenvalues. This significantly improves the training stability and convergence speed of the model in strongly constrained multi-task optimization scenarios, effectively solving the optimization ill-conditioning problem caused by gradient imbalance.

[0042] Furthermore, the gradient spectrum feature value corresponding to each loss component is calculated in real time based on the neural tangent, and the weight coefficient is dynamically adjusted according to the gradient spectrum feature value. Each loss component refers to the data fitting loss, physical residual loss and mixed boundary constraint loss mentioned above. The specific adjustment steps for dynamically adjusting the weight coefficients based on the gradient spectrum feature values ​​are as follows: when the gradient spectrum feature value of the loss component of the mixed boundary constraint term is detected to be lower than a preset threshold, the corresponding mixed boundary constraint weight coefficient is adaptively increased. Specifically: Let the gradient spectrum eigenvalue of the boundary constraint term in the t-th training iteration be... The preset threshold is ,when When the boundary constraint weight coefficients are updated, they are updated as follows:

[0043] Where η is the weight update step size, ε is the stability term, λmax is the upper limit of the boundary constraint weights, and λbound is the boundary constraint weight coefficient. This method allows the optimization priority of the boundary constraint terms during training to adaptively adjust with changes in their gradient spectrum eigenvalues, thereby improving the model's satisfaction of physical constraints such as obstacle boundaries and depositional boundaries.

[0044] Specifically: when a significant decay in the gradient contribution of the boundary constraint term is detected, λ is adaptively increased. b1 , λ b2 By artificially constructing high-potential-energy physical barriers, the optimizer is forced to prioritize satisfying the physical conservation conditions at complex boundaries, thereby enhancing the numerical stability of the model under strongly nonconvex topological conditions.

[0045] Furthermore, the acquisition of the three-dimensional meteorological field data in step S1 includes: using a multi-scale nested numerical simulation method to nest the regional-scale background wind field with the local micro-scale flow field, wherein the immersion boundary method is used to represent the non-convex geometric boundary formed by bridge piers and / or shoreline building complexes.

[0046] The multi-scale nested numerical simulation method includes the following steps: A WRF-OpenFOAM unidirectional coupling strategy is adopted, using the WRF model to simulate the background wind field and thermal structure of the research area at a kilometer-level resolution, serving as the boundary driving condition for the OpenFOAM micro-scale simulation; at the waterway scale, a 100-meter-level resolution partial analytical simulation of turbulence is employed (WRF-LES gray zone method), focusing on obtaining the dominant flow pattern and turbulence statistical characteristics of the near-surface layer of the inland waterway; in bridge areas and densely built-up areas, locally intensified nested domains are set up, and for the strongly non-convex geometric boundaries formed by the building complexes and bridge structures on both sides of the inland river, the immersion boundary method is introduced. The method (IBM) introduces a volumetric force source term into the momentum equation to represent the solid structure equivalently, avoiding the difficulties in generating body-fitted meshes and numerical instability under complex topological conditions. Through multi-scale nested simulation, it focuses on characterizing key micrometeorological features such as canyon winds and backflow zones induced by coastal building clusters, local turbulence enhancement and shear layer structures induced by bridge piers and beams, and the water surface canopy effect formed by the thermal difference between the water surface and the land surface. These features serve as a unified meteorological reference field for subsequent diffusion simulations and data-driven model training.

[0047] Furthermore, the specific process of acquiring ship emission source data in step S1 is as follows: ship trajectory and operating condition data are collected through the Automatic Identification System (AIS), the instantaneous emission intensity of the ship is inverted, and the Lagrange-Euler hybrid description method is adopted. A Gaussian kernel function is introduced to map the discrete moving point source into a continuous Euler grid three-dimensional dynamic emission source term field Q(x,y,z,t). The neural network model described in step S2 is a fully connected deep neural network. The input of the neural network model includes ship emission source data and meteorological environmental data. The ship emission source data includes the spatiotemporal coordinates of the ship emission source and the ship source term. The meteorological environmental data includes wind speed, wind direction, atmospheric stability, rainfall intensity and / or turbulence diffusion parameters extracted or interpolated from three-dimensional meteorological field data. The neural network model outputs pollutant concentration. By calculating each predicted point in the three-dimensional spatial grid of the measured section, the reconstructed ship exhaust gas diffusion concentration field is obtained. Furthermore, the forward inference speed in step S5 is on the order of milliseconds to seconds; the ship exhaust gas diffusion field reconstruction method based on physical information fusion also includes the following step: transmitting the reconstructed concentration field results to the monitoring terminal for visualization.

[0048] A rapid reconstruction system for ship exhaust gas diffusion field based on physical information fusion includes: a shore-based Internet of Things (IoT) system, a monitoring terminal, a processor, and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the ship exhaust gas diffusion field reconstruction method based on physical information fusion is implemented. The shore-based IoT system is connected to the processor and is used to collect multi-source input data and transmit it to the processor. The monitoring terminal is connected to the processor and is used to display the output results of the processor implementing the ship exhaust gas diffusion field reconstruction method based on physical information fusion.

[0049] The working principle of this invention: The ship exhaust gas diffusion field reconstruction method based on physical information fusion of this invention includes the following steps: Step 1: Acquire multi-source input data through the shore-based IoT system.

[0050] (1) Meteorological data acquisition: The three-dimensional wind speed field u(x,y,z,t), atmospheric temperature, humidity and rainfall data of the research section are collected in real time by meteorological sensor nodes (including anemometers, temperature and humidity sensors, rain gauges, etc.) deployed at shore-based monitoring stations. When necessary, the data can be supplemented by combining the WRF-OpenFOAM multi-scale nested meteorological field numerical simulation system. The background wind field at the regional scale (kilometer-level resolution) and the flow field at the local micro scale (meter-level resolution) are nested through a one-way coupling strategy. For the strong non-convex geometric boundaries formed by physical structures such as bridge piers and shoreline building complexes, the immersion boundary method is used for equivalent expression, which effectively characterizes the complex flow phenomena such as canyon wind, backflow zone and local vortex induced by physical structures. (2) Acquisition of ship emission source data: Ship trajectory and operating condition data are collected by AIS receiver to invert the instantaneous emission intensity of ships; the Lagrange-Euler hybrid description method is adopted, and the Gaussian kernel function is introduced to map the discrete moving point sources of N ships into a continuous Euler grid three-dimensional dynamic emission source term field Q(x,y,z,t); (3) Acquisition of measured concentration data: Sparse measured pollutant concentration data are collected by shore-based fixed monitoring IoT nodes (including multi-point sensor arrays mounted on the bridge structure) for physical constraint fine-tuning and model verification.

[0051] Step 2: Construct a neural network model with deep embedding of physical mechanisms (core architecture of boundary adaptive physical information neural network).

[0052] (1) The neural network backbone mounted on the edge computing processor adopts a fully connected deep neural network structure. The input is three-dimensional spatiotemporal coordinates (x,y,z,t) and environmental parameters (wind speed, wind direction, atmospheric stability, rainfall intensity, etc.), and the output is the predicted pollutant concentration field c(x,y,z,t). (2) Using automatic differentiation technology, partial derivatives of the network output concentration field c are obtained, and physical residual terms based on the following modified convection-diffusion-reaction equation are constructed. The wet deposition mechanism is explicitly introduced to correct the nonlinear attenuation under rain and fog weather:

[0053] In the formula, For dynamic ship source terms reconstructed based on AIS, This represents the pollutant concentration field predicted by a neural network. This is a three-dimensional wind field vector; is the turbulent diffusion coefficient, which characterizes the ability of atmospheric turbulence to mix pollutants. The comprehensive settlement coefficient taking into account the effects of rain / fog.

[0054] Step 3: Design a block-mixed boundary constraint mechanism under non-convex topology conditions.

[0055] The "boundary" involved in this step specifically refers to the solid-fluid interface formed by the mechanical obstruction and flow-around effects of concrete physical structures such as bridge piers, shoreline buildings, and bridge box girders on exhaust plumes in inland waterways, rather than the general virtual boundary or free flow field boundary of the fluid dynamics computational domain. Given the complex geometric and topological characteristics of inland waterways, the single boundary condition assumption in traditional PINNs is insufficient to simultaneously describe the processes of structure obstruction and surface subsidence. Therefore, a block-based hybrid boundary constraint mechanism is constructed, explicitly introducing physical obstacle boundaries during network training. Settlement boundary with underlying surface Physical constraints.

[0056] To address the two distinctly different types of physical boundaries, a segmented hybrid boundary constraint mechanism is constructed: Neumann-type flux-free constraints are applied to the surfaces of shoreline building complexes and bridge piers. This forces pollutant streamlines to bypass and be blocked, preventing the model from producing non-physical diffusion results that penetrate buildings. For both water and land surfaces, Robin-type mixed boundary conditions are applied. ,in It is a dry / wet deposition velocity parameter used to characterize the flux of pollutants being adsorbed and removed by water bodies or land surfaces, effectively alleviating the problem of systematic overestimation of near-surface concentration.

[0057] Step 4: Adaptive multi-task weight regulation based on the Neural Tangent Kernel.

[0058] To address the ill-conditioned optimization problem caused by the imbalance of gradient magnitudes among the physical residual, data observation, and mixed boundary constraint terms during training, an adaptive weighting strategy using a Neural Tangent Kernel (NTK) is introduced to dynamically adjust the weighting of the multi-task loss function. The constructed multi-task joint loss function is shown in Equation (2):

[0059] In the formula, This represents the fitting loss of sparse observation data, which comes from fixed-point sampling of a high-fidelity simulated field or actual measured values. , These are the weighting coefficients for the sparse observation data fitting term and the physical residual term, respectively; n obs n is the unit outward normal vector of the boundary surface of the solid obstacle; dep is the unit outward normal vector of the settlement boundary surface of the underlying surface; , These are adaptive weights for physical obstacles and underlying surface settlement boundaries, respectively. During training iterations, NTK is used to calculate the gradient spectrum feature values ​​corresponding to each loss component in real time. When a significant decay in the gradient contribution of the boundary constraint term is detected, the weights are adaptively increased. , By artificially constructing high-potential-energy physical barriers, the optimizer is forced to prioritize satisfying the physical conservation conditions at complex boundaries, thereby enhancing the numerical stability of the model under strongly nonconvex topological conditions.

[0060] Step 5: Model training, rapid inference, and supervisory output.

[0061] (1) Training: The Adam / L-BFGS hybrid optimization strategy was adopted, and the multi-scenario high-fidelity numerical simulation dataset was used for pre-training. The sparse measured concentration data collected by the shore-based IoT nodes in step one was used for physical constraint fine-tuning. (2) Fast reasoning: The trained boundary adaptive physical information neural network is run on the shore-based edge computing device equipped with a processor. For the meteorological and emission inputs at the new monitoring time, millisecond- to second-level forward reasoning is performed to reconstruct the three-dimensional concentration field c(x,y,z,t); (3) Regulatory output: The reconstructed concentration field results are transmitted to the green shipping regulatory screen for visualization, so as to realize precise and intelligent supervision of exhaust emissions from inland waterway vessels.

[0062] Block-based hybrid boundary constraint mechanism For complex non-convex topological boundaries of inland waterways, this paper organically combines Neumann-type flux-free constraints on the surfaces of buildings / bridges with Robin-type hybrid settlement constraints on the water / land surface. The paper achieves accurate embedding of block-based hybrid boundary physical constraints within a physical information neural network training framework, fundamentally solving the problem of non-physical penetration predictions caused by single boundary assumptions in complex topological environments in existing physical information neural network methods.

[0063] Multi-task adaptive weight control strategy based on neural tangent nucleus To address the imbalance of gradient magnitudes among physical residuals, data observations, and various types of boundary constraints during collaborative training, this paper proposes to utilize the eigenvalues ​​of the neural tangent kernel to evaluate the gradient contribution of each loss component in real time, and dynamically adjust the weight coefficients based on the proportion of eigenvalues. This significantly improves the training stability and convergence speed of the model in strongly constrained multi-task optimization scenarios, effectively solving the ill-conditioned optimization problem caused by gradient imbalance.

[0064] Physical constraints of the modified convection-diffusion-reaction equation considering wet deposition mechanism By explicitly incorporating the comprehensive wet deposition coefficient λd, which is related to rainfall and relative humidity, into the physical control equation (convection-diffusion-reaction equation), accurate modeling of the nonlinear decay process of pollutants under special meteorological conditions such as rain and fog is achieved. This effectively solves the problem of systematic overestimation of concentration under rainfall conditions by traditional methods and broadens the applicability of physical information neural networks under various meteorological conditions.

[0065] Characterization of Ship Dynamic Emission Sources Based on the Lagrange-Euler Mixture Method The Lagrange moving point source trajectory driven by the Automatic Identification System (AIS) data is mapped into a continuous Eulerian grid source term field through a Gaussian kernel function smoothing operator. This achieves a physically consistent and continuous representation of the moving emission source within a fixed computational domain, avoids numerical oscillations caused by discrete moving point sources, and supports a refined characterization of the superimposed emission effects of multiple ships.

[0066] Physical information embedding method: Utilize automatic differentiation technology (frameworks such as PyTorch / TensorFlow) to calculate the residuals of the physical control equations in real time, and jointly optimize them with boundary constraints and data fitting to ensure global physical consistency; Complex boundary geometry modeling: Describes the non-convex topological boundaries (bridge piers, building complexes, etc.) of inland waterways through signed distance functions or direct geometric constraint point sets, supporting arbitrary complex boundary shapes; Multi-scale meteorological field driven: High-resolution three-dimensional meteorological fields are provided by multi-scale nested numerical simulation (WRF-OpenFOAM / Large Eddy Simulation) as physical constraint inputs to ensure the physical accuracy of meteorological parameters in the diffusion equation; Fast inference mechanism: After the model is trained, it only requires millisecond to second-level neural network forward inference for new scenarios (without resolving partial differential equations), which is 2 to 3 orders of magnitude faster than numerical simulation, and realizes real-time diffusion field reconstruction; Application scenarios: Primarily targeting the rapid reconstruction of ship exhaust emission diffusion in inland waterways (including complex boundaries such as bridges and shoreline structures), it can also be extended to port areas and other atmospheric diffusion scenarios with complex boundary constraints.

[0067] Example 2 Based on Example 1, taking a cross-river bridge area in the Wuhan section of the Yangtze River as a typical scenario, a boundary adaptive physical information neural network diffusion field reconstruction method is implemented using a shore-based IoT monitoring system deployed on the bridge structure (including 6 NOx / SO2 sensor nodes, 1 AIS receiver, 1 meteorological station, and 1 edge computing server). The steps are as follows: (I) Multi-source data acquisition and processing (system hardware layer) 1. Meteorological data: Real-time data on wind speed and direction at a height of 10 meters, temperature at 2 meters, surface air pressure and precipitation are collected by shore-based meteorological sensor nodes. The data are combined with the European Reanalysis Dataset (ERA5) to drive the WRF regional meteorological model, providing initial and boundary conditions for the numerical simulation of the three-dimensional meteorological field. 2. Ship emission source data: Real-time ship trajectories (ship speed, heading, latitude and longitude, etc.) in the Wuhan section of the Yangtze River are collected by AIS receivers. Combined with the inland waterway ship main engine load mapping model that takes into account the influence of water flow, the instantaneous NOx and SO2 emission intensity of ships (unit: g / s) is inverted to construct a three-dimensional time-varying emission source term field Q(x,y,z,t); 3. Measured concentration data: The time series of NOx and SO2 concentrations were collected by 6 fixed sensor nodes mounted on the bridge when ships passed by. These data were used as sparse observation data for model fine-tuning and verification.

[0068] (II) Construction of a three-dimensional meteorological field (numerical simulation layer) 1. Regional scale (3 km resolution): Run WRF mode to output hourly 3D background wind field; 2. Channel scale (100-meter resolution): The large eddy simulation method is used to obtain the statistical characteristics of near-surface turbulence, focusing on simulating key micro-meteorological structures such as canyon winds and backflow zones; 3. Bridge area microscale (resolution 5-10 meters): The immersion boundary method is used in OpenFOAM to finely characterize the solid structures such as bridge piers (rectangular cross-section, about 8 meters wide), bridge box girder, and building complexes on both banks (about 30-100 meters high), simulating the wake vortex system induced by solid obstacles and the shear layer enhancement structure; the three-dimensional wind speed field u(x,y,z,t) and turbulence diffusion coefficient tensor K(x,y,z,t) of the bridge area are obtained as meteorological driving inputs for the boundary adaptive physical information neural network.

[0069] (III) Geometric Modeling of Entity Boundaries The three-dimensional geometric coordinates of bridge piers and shoreline building complexes in the study area are extracted to construct constraint point sets for two types of physical boundaries: Ω_Obstacles: Solid-fluid contact surfaces such as bridge pier surfaces and building exterior walls (physical meaning: exhaust plumes are mechanically blocked by bridge piers, forming a stagnant high-pressure zone on the windward side of the pier and a negative pressure wake zone on the leeward side, resulting in a strong flow bypass effect). Ω_Underlying surface: Water surface and land surface (physical meaning: pollutants settle onto the water surface / land surface due to gravity, rain leaching, etc.).

[0070] (iv) Construction and training of boundary adaptive physical information neural network model 1. Network architecture: An 8-layer fully connected deep neural network is used, with 256 neurons in each layer. The activation function is the hyperbolic tangent function. The input dimension is 5 (x, y, z, t and meteorological condition encoding), and the output is the concentration prediction value c(x, y, z, t). 2. Physical residuals (corrected three-dimensional convection-diffusion-reaction equations): As shown in formula (1); 3. Entity boundary constraints: Bridge piers / building surfaces (Ω_obstacles): c / n=0 (no flux, simulating physical mechanical obstruction) Water surface / land surface (Ω_underlying surface): c / n = -vd·c (settlement, vd = 0.01 m / s on sunny days, vd ≈ 0.05 m / s during rainfall) 4. Multi-task joint loss and adaptive weights of the neural tangent kernel: As shown in formula (2); The gradient feature value of the neural tangent kernel is calculated every 100 training steps. If the gradient feature value of the obstacle constraint or the underlying surface constraint is lower than 50% of the mean, the corresponding weight is automatically increased to ensure that the physical constraints of the physical boundaries of the bridge piers and other entities remain effective throughout the training process. 5. Optimization strategy: Train the Adam optimizer for 20,000 steps (initial learning rate 0.001), then switch to L-BFGS for fine optimization for 5,000 steps until convergence.

[0071] (v) Output of rapid reasoning and monitoring system After the model training is completed, it is deployed on a shore-based edge computing server (equipped with a GPU accelerator card) in the bridge area: Whenever new weather and AIS data arrive, the edge computing device automatically triggers forward inference; The reconstructed three-dimensional NOx / SO2 concentration field c(x,y,z,t) is transmitted to the green shipping supervision screen to display the exhaust gas diffusion situation in the bridge area in real time using a pseudo-color image. This invention uses a fully connected deep neural network as its backbone, but it can also be replaced by the following architecture: Fourier Feature Networks: By encoding the input coordinates with random Fourier features at high frequency, Fourier Feature Networks can effectively alleviate the spectral bias problem of physical information neural networks in high-frequency oscillating concentration fields and are suitable for near-source regions with sharp concentration gradients. DeepONet or neural operator architecture: It can directly learn the operator mapping from meteorological parameters to concentration fields, and has a stronger generalization advantage in fast reasoning in multiple scenarios; Alternatives to boundary constraint types: Dirichlet-type boundary conditions (directly specifying the boundary concentration as zero) can replace Neumann-type flux-free constraints, simplifying the boundary processing logic within acceptable computational accuracy ranges. Penalty function method to replace hard constraints: Boundary constraints are added to the loss function in the form of soft constraint penalty terms, which can simplify the workload of geometric modeling when it is difficult to extract the boundary of complex surfaces. Signed distance function constraints: Boundary conditions are precisely embedded using a signed distance function, which is especially suitable for the accurate description of complex 3D geometric boundaries.

[0072] Alternatives to emission source term characterization: Using the Gaussian plume model of chimney emissions directly as the initial source term approximation is suitable for rapid application scenarios where detailed meteorological simulation data is lacking. The fusion of column concentration data based on satellite remote sensing and ground observation can replace the emission source items retrieved by the Automatic Identification System for ships, and is applicable to marine and large-scale water scenarios. Based on the measured emission data from the Portable Emission Measurement System (PEMS), the source item is directly input for high-precision scenario verification.

[0073] Alternative methods for meteorological field input: The measured meteorological station data (wind speed, wind direction, etc.) can be directly used to replace the numerical simulation meteorological field, and it has good applicability in port areas where meteorological stations are densely deployed. Satellite wind field remote sensing data (such as SAR) can be used as a supplement to meteorological field input, which is especially suitable for coastal / estuary areas.

[0074] In summary, this invention relies on a shore-based edge computing device equipped with a processor. It receives measured concentration data collected by shore-based monitoring IoT nodes, ship trajectory and operating condition data output by the Automatic Identification System (AIS) receiver, and real-time meteorological data collected by meteorological sensors. It then calculates and outputs a reconstructed diffusion concentration field, which can be further transmitted to green shipping supervision and high-emission area monitoring, achieving precise monitoring and control of ship exhaust diffusion under complex micro-meteorological conditions in inland waterways. This invention proposes a boundary adaptive weighted physical information neural network (hereinafter referred to as "boundary adaptive physical information neural network") for rapid reconstruction of ship exhaust diffusion fields. Through a three-layer mechanism of "deep embedding of physical mechanisms—block-based hybrid boundary constraints—adaptive control of neural tangent kernel gradient," it solves the core technical problems of existing methods, such as poor physical consistency, unstable training, and insufficient real-time performance in complex physical boundary environments in inland waterways, achieving stable and rapid reconstruction of ship exhaust diffusion fields in complex and restricted waterways.

[0075] ① Effectively solved the engineering and technical problem of the difficulty in accurately modeling the mechanical obstruction effect of physical obstacles such as bridge piers and shoreline building complexes on exhaust gas diffusion in inland waterways; ② Physical consistency: The convection-diffusion-reaction equation and the boundary conditions of the solid structure are embedded into the deep learning framework to ensure that the predicted concentration satisfies the physical conservation law; ③ Computational efficiency: The inference speed is 2 to 3 orders of magnitude faster than traditional numerical simulation, and it can run in real time on shore-based edge computing devices; ④ Adaptability to multiple meteorological conditions: The wet deposition mechanism is explicitly introduced, and the prediction accuracy is still high under special meteorological conditions such as rainfall and haze; ⑤ Training stability: The adaptive weighting strategy of the neural tangent kernel effectively prevents the entity boundary constraints from failing during training, ensuring the physical correctness of complex boundary regions; ⑥ System closed loop: It achieves data closed loop with shore-based IoT monitoring system and ship supervision screen, and has complete engineering application value.

[0076] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0077] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for reconstructing the ship exhaust gas diffusion field based on physical information fusion, characterized in that: Includes the following steps: Acquire multi-source input data for the target inland waterway area through a shore-based Internet of Things (IoT) system; The multi-source input data includes: three-dimensional meteorological field data, ship emission source data, and measured pollutant concentration data detected by shore-based sensors; Physical residuals are constructed based on three-dimensional convection-diffusion-reaction equations, and the physical residuals are embedded into a deep neural network training framework using automatic differentiation techniques. During neural network training, physical constraints of physical obstacle boundaries and underlying surface settlement boundaries are explicitly introduced to construct a block-based hybrid boundary constraint mechanism. Based on sparse observation data fitting loss, physical residual loss, and hybrid boundary constraint loss, an adaptive weighting strategy of neural tangent kernel is introduced to construct a multi-task joint loss function and train the neural network model. The trained neural network performs forward inference based on real-time input three-dimensional meteorological field data and ship emission source data to reconstruct the ship exhaust gas diffusion concentration field.

2. The ship exhaust gas diffusion field reconstruction method based on physical information fusion as described in claim 1, characterized in that: The physical residuals of the three-dimensional convection-diffusion-reaction equations are: Where S(X,t) is the dynamic ship source term reconstructed based on AIS, X is the ship position vector, and t is time; The pollutant concentration field predicted by the neural network. denoted as the gradient of the pollutant concentration field predicted by the neural network; u is the three-dimensional wind field vector; D is the turbulent diffusion coefficient, which characterizes the mixing ability of atmospheric turbulence on pollutants; k dep The comprehensive settlement coefficient taking into account the effects of rain / fog.

3. The ship exhaust gas diffusion field reconstruction method based on physical information fusion as described in claim 1, characterized in that: The physical constraints of the solid obstacle boundary are: applying a flux-free constraint to the surface of the solid obstacle in the inland waterway; the physical constraints of the underlying surface settlement boundary are: applying a settlement constraint to the water surface and / or land surface. The physical obstacles include bridge piers and / or shoreline structures. Neumann-type flux-free constraints are applied to the surfaces of these physical obstacles in inland waterways. ; This is used to represent the flow around and obstruction of pollutants, avoiding the model from producing non-physical diffusion results that penetrate buildings. obs Let be the unit outward normal vector of the boundary surface of the solid obstacle; Apply Robin-type mixed boundary conditions to both water and land surfaces. ,in Here are the dry / wet settling velocity parameters, D is the turbulent diffusion coefficient, and n is the turbulent diffusion coefficient. dep Let be the unit outward normal vector of the settlement boundary surface of the underlying surface. The pollutant concentration field predicted by the neural network. This represents the gradient of the pollutant concentration field predicted by the neural network.

4. The ship exhaust gas diffusion field reconstruction method based on physical information fusion as described in claim 1, characterized in that: The multi-task joint loss function is: in, R represents the fitting loss for sparse observation data; PDE For physical residuals; These are the weighting coefficients for the fitting term of the sparse observation data; n represents the weighting coefficient of the physical residual term; obs n is the unit outward normal vector of the boundary surface of the solid obstacle; dep λ is the unit outward normal vector of the underlying surface at the settlement boundary; b1 This refers to the adaptive weighting coefficient of the entity obstacle boundary; λ b2 This refers to the adaptive weighting coefficient of the underlying surface settlement boundary; D The turbulent diffusion coefficient; The pollutant concentration field predicted by the neural network; This refers to the gradient of the pollutant concentration field predicted by the neural network; These are the dry / wet settling velocity parameters; Indicates the boundary of a physical obstacle; This indicates the settlement boundary of the underlying surface.

5. The ship exhaust gas diffusion field reconstruction method based on physical information fusion as described in claim 4, characterized in that: During the training iteration, the gradient spectrum feature values ​​corresponding to each loss component are calculated in real time based on the neural tangent, and the weight coefficients λ are dynamically adjusted according to the gradient spectrum feature values. b1 and / or λ b2 .

6. The ship exhaust gas diffusion field reconstruction method based on physical information fusion as described in claim 5, characterized in that: The gradient spectrum feature value corresponding to each loss component is calculated in real time based on the neural tangent, and the weight coefficient is dynamically adjusted according to the gradient spectrum feature value. Each loss component refers to the data fitting loss, physical residual loss and mixed boundary constraint loss mentioned above. The specific adjustment steps for dynamically adjusting the weight coefficients based on the gradient spectrum feature values ​​are as follows: when the gradient spectrum feature value of the mixed boundary constraint loss component is detected to be lower than a preset threshold, the corresponding mixed boundary constraint weight coefficient is adaptively increased. Specifically: Let the gradient spectrum eigenvalue of the boundary constraint term in the t-th training iteration be... The preset threshold is ,when When the boundary constraint weight coefficients are updated, they are updated as follows: Where η is the weight update step size, ε is the stability term, λmax is the upper limit of the boundary constraint weights, and λbound is the boundary constraint weight coefficient.

7. The ship exhaust gas diffusion field reconstruction method based on physical information fusion as described in claim 1, characterized in that: The acquisition of the three-dimensional meteorological field data includes the following steps: using a multi-scale nested numerical simulation method, the regional scale background wind field and the local micro-scale flow field are nested, wherein the immersion boundary method is used to represent the non-convex geometric boundary formed by bridge piers and / or shoreline building complexes. The multi-scale nested numerical simulation method includes a one-way coupling strategy of mesoscale meteorological models and computational fluid dynamics models to obtain background wind fields and thermal structures at kilometer-level resolution, which serve as boundary driving conditions for microscale simulations. At the waterway scale, a partial analytical simulation method for turbulence with a resolution of 100 meters was used to obtain the dominant flow pattern and turbulence statistical characteristics of the near-surface layer of inland waterways. Locally encrypted nested domains are set up in bridge areas and densely built areas. For non-convex geometric boundaries formed by building groups and bridge structures, the immersion boundary method is adopted. The volume force source term is introduced into the momentum equation to replace the body-fit solution of the solid structure. The specific process for acquiring ship emission source data is as follows: ship trajectory and operating condition data are collected through the Automatic Identification System (AIS), the instantaneous emission intensity of the ship is inverted, and the Lagrange-Euler hybrid description method is adopted. A Gaussian kernel function is introduced to map discrete moving point sources into a continuous Euler grid three-dimensional dynamic emission source term field.

8. The ship exhaust gas diffusion field reconstruction method based on physical information fusion as described in claim 1, characterized in that: The neural network model can be any one of a fully connected deep neural network, a Fourier feature network, or a deep operator network. The input to the neural network model includes ship emission source data and meteorological environmental data; wherein, the ship emission source data includes the spatiotemporal coordinates of the ship emission source and the ship source term, and the meteorological environmental data includes wind speed, wind direction, atmospheric stability, rainfall intensity and / or turbulence diffusion parameters extracted or interpolated from three-dimensional meteorological field data; The neural network model outputs pollutant concentration, and the reconstructed ship exhaust gas diffusion concentration field is obtained by calculating each predicted point in the three-dimensional spatial grid of the measured section.

9. The ship exhaust gas diffusion field reconstruction method based on physical information fusion as described in claim 1, characterized in that: The aforementioned method for reconstructing the ship exhaust gas diffusion field based on physical information fusion also includes the following steps: transmitting the reconstructed concentration field results to a monitoring terminal for visualization.

10. A rapid reconstruction system for ship exhaust gas diffusion field driven by physical information fusion, characterized in that: include: The system comprises a shore-based Internet of Things (IoT) system, a monitoring terminal, a processor, and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the ship exhaust gas diffusion field reconstruction method based on physical information fusion as described in any one of claims 1-9. The shore-based IoT system is connected to the processor and is used to collect multi-source input data and transmit it to the processor. The monitoring terminal is connected to the processor and is used to display the output results of the processor implementing the ship exhaust gas diffusion field reconstruction method based on physical information fusion.

Images