A method and system for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision.
By fusing physical information neural networks with multi-plane vision and combining fluid dynamics control equations, the problems of data loss and dynamic evolution of physical parameters in gas-liquid two-phase release monitoring are solved, achieving high-precision and robust three-dimensional physical field reconstruction, which is applicable to the fields of petrochemical and nuclear power safety.
Patent Information
- Application Number
- CN202611427707.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-15
- Publication Date
- 2026-11-17
AI Technical Summary
Existing gas-liquid two-phase leakage monitoring technologies have significant shortcomings in terms of physical field reconstruction consistency, robustness under complex operating conditions, and occlusion compensation capabilities, making it difficult to meet the industrial requirements for high-precision reconstruction and adaptive identification of physical property parameters.
By employing a method based on physical information neural networks and multi-plane vision fusion, a composite loss function optimization model is constructed using multi-plane vision observation data and process parameters, combined with fluid dynamics control equations, to achieve real-time reconstruction of the three-dimensional physical field, thus solving the problems of missing visual observation data and dynamic evolution of physical property parameters.
It improves the physical consistency, occlusion compensation capability, and adaptive identification capability of physical property parameters in the reconstruction of the physical field of gas-liquid two-phase discharge, enhances robustness and accuracy under complex working conditions, and reduces dependence on hardware and data acquisition.
Smart Images

Figure CN123118918A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of petrochemical safety monitoring technology, and in particular to a method and system for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision. Background Technology
[0002] Two-phase gas-liquid venting processes are widely used in petrochemical, energy, and nuclear power safety fields, and are a key link in ensuring the intrinsic safety of high-pressure reactors and process systems. Compared to single-phase fluids, two-phase gas-liquid venting fluids have complex characteristics such as unsteadiness, multiphase coupling, and rapid evolution. During venting, the sudden pressure drop induces flash evaporation, leading to drastic phase changes and fluctuations in component density. The velocity gradient, pressure distribution, and phase ratio within the flow field exhibit extremely strong inhomogeneities. Therefore, accurately monitoring the venting volume and assessing its safety impact on the environment and equipment has become a core technological bottleneck restricting the improvement of chemical safety assurance capabilities.
[0003] Currently, industrial monitoring of gas-liquid two-phase discharge mainly relies on traditional flow meters and digital image processing technology. However, these technologies still have significant shortcomings in practical applications. Regarding measurement principles, traditional differential pressure or Coriolis flow meters are mostly designed based on single-phase flow or fixed-proportion mixed flow models, making it difficult to adapt to the drastic changes in phase ratios during discharge. This can easily lead to large measurement deviations under high-pressure flash evaporation conditions. In terms of visual monitoring, related technologies mostly employ two-dimensional image processing based on high-speed cameras or binocular velocimetry methods. The visual information from a single plane cannot characterize the continuous flow field in three-dimensional space. Furthermore, the discharged fluid typically has high concentrations and complex structures, easily causing visual overlap and line-of-sight obstruction, resulting in data loss in key areas within the flow field. In addition, related image recognition methods mostly remain at the data fitting level, without incorporating fluid dynamics and physical constraints, failing to truly reflect the physical nature of the discharge process. Regarding system robustness, related monitoring schemes are mostly based on discrete post-processing modes, lacking adaptive identification capabilities for fluid properties. When operating conditions fluctuate drastically or flow regimes change, monitoring models with preset fixed parameters are prone to failure due to physical mismatch.
[0004] In summary, existing gas-liquid two-phase leakage monitoring technologies have significant shortcomings in three core dimensions: consistency of physical field reconstruction, robustness under complex operating conditions, and occlusion compensation capability. They are unable to simultaneously meet the industrial requirements of high-precision reconstruction and adaptive identification of physical property parameters.
[0005] Therefore, there is an urgent need to provide a real-time reconstruction method for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision, in order to solve the above problems. Summary of the Invention
[0006] The purpose of this application is to provide a real-time reconstruction method and system for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision, so as to solve the problems of lack of physical constraints, data loss due to line of sight obstruction, and difficulty in identifying the dynamic evolution of physical property parameters under flash evaporation conditions in the existing gas-liquid two-phase discharge monitoring technology, and improve the physical consistency, obstruction compensation capability, adaptive identification capability of physical property parameters, and robustness under complex working conditions of the reconstructed physical field of gas-liquid two-phase discharge.
[0007] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision, the method comprising: The system acquires multi-plane visual observation data and process parameters of the discharge source within the observation area to be reconstructed in real time, and selects multiple physical points within the observation area to be reconstructed; the multi-plane visual observation data includes droplet dynamic sequence images of multiple discrete observation planes within the observation area to be reconstructed, acquired through the multi-plane visual acquisition system. Preprocessing of multi-plane visual observation data yields the spatiotemporal coordinates of each visual observation point within the observation area to be reconstructed, as well as the observed values of droplet velocity. The spatiotemporal coordinates of each visual observation point and each physical coordinate point in the observation area to be reconstructed are input into the initial gas-liquid two-phase discharge physical field reconstruction model, which outputs the initial predicted values of the physical field variables of each visual observation point and each physical coordinate point in the observation area to be reconstructed. The initial gas-liquid two-phase discharge physical field reconstruction model is obtained by training a physical information neural network using a training set. The physical field variables include: three-dimensional velocity vector, pressure, density, and dynamic viscosity. Based on the initial predicted values of the physical field variables at each visual observation point, the initial predicted values of the physical field variables at each physical alignment point, the observed values of the droplet velocity, and the process parameters at the discharge source, a composite loss function is constructed. The composite loss function includes: data observation loss, physical residual loss, initial condition loss, and boundary condition loss. The network parameters of the initial gas-liquid two-phase venting physical field reconstruction model are optimized with the composite loss function as the objective until the composite loss function converges, thus obtaining the target gas-liquid two-phase venting physical field reconstruction model. The spatiotemporal coordinates of the observation area to be reconstructed are input into the target gas-liquid two-phase discharge physical field reconstruction model, and the predicted value of the three-dimensional physical field of the observation area to be reconstructed is output.
[0008] Secondly, this application provides a real-time reconstruction system for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision. This system is used to implement the aforementioned real-time reconstruction method for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision. The real-time reconstruction system includes: The data acquisition and physical point selection unit is used to acquire multi-plane visual observation data and process parameters of the discharge source in the observation area to be reconstructed in real time, and select multiple physical points in the observation area to be reconstructed; the multi-plane visual observation data includes droplet dynamic sequence images of multiple discrete observation planes in the observation area to be reconstructed, acquired by the multi-plane visual acquisition system. The preprocessing unit is used to preprocess the multi-plane visual observation data to obtain the spatiotemporal coordinates of each visual observation point in the observation area to be reconstructed and the observed values of the droplet motion velocity. The initial reconstruction unit is used to input the spatiotemporal coordinates of each visual observation point and each physical coordinate point in the observation area to be reconstructed into the initial gas-liquid two-phase discharge physical field reconstruction model, and output the initial predicted values of the physical field variables of each visual observation point and each physical coordinate point in the observation area to be reconstructed. The initial gas-liquid two-phase discharge physical field reconstruction model is obtained by training a physical information neural network using a training set. The physical field variables include: three-dimensional velocity vector, pressure, density, and dynamic viscosity. The loss function construction unit is used to construct a composite loss function based on the initial predicted values of the physical field variables at each visual observation point, the initial predicted values of the physical field variables at each physical alignment point, the observed values of the droplet velocity, and the process parameters at the discharge source. The composite loss function includes: data observation loss, physical residual loss, initial condition loss, and boundary condition loss. The model optimization unit is used to optimize the network parameters of the initial gas-liquid two-phase discharge physical field reconstruction model with the composite loss function as the objective until the composite loss function converges, thus obtaining the target gas-liquid two-phase discharge physical field reconstruction model. The target reconstruction unit is used to input the spatiotemporal coordinates of the observation area to be reconstructed into the target gas-liquid two-phase discharge physical field reconstruction model, and output the predicted value of the three-dimensional physical field of the observation area to be reconstructed.
[0009] According to the specific embodiments provided in this application, this application has the following technical effects: This application discloses a real-time reconstruction method and system for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision. By inputting the spatiotemporal coordinates of each visual observation point and each physical coordinate point into an initial physical information neural network model, initial predicted values of the physical field variables are obtained. Based on this, a data observation loss is constructed based on the deviation between the predicted three-dimensional velocity vector values and the observed droplet velocity values at each visual observation point. A physical residual loss is constructed based on the residuals obtained by substituting the predicted values at each physical coordinate point into the mass conservation equation and the momentum conservation equation. Furthermore, initial condition loss and boundary condition loss are constructed in conjunction with process parameters to form a composite loss function for model optimization. This embeds the fluid dynamics control equations into the reconstruction process, solving the problems of lack of physical constraints in pure data fitting and difficulty in satisfying mass and momentum conservation in the reconstructed flow field, thus improving physical consistency. Simultaneously, the physical field... The physical alignment point covers the entire observation area to be reconstructed, independent of visual visibility. This allows for the consistent completion of data in areas where visual observation is missing due to droplet concentration, overlap, or occlusion by physical residual constraints. This solves the problems of insufficient visual information from a single plane and data loss in key areas caused by line-of-sight occlusion, improving occlusion compensation capabilities. Furthermore, the physical field variables output by the model include density and dynamic viscosity. A composite loss function is constructed using the visual observation values under the current operating conditions, the predicted values of the physical alignment point, and the process parameters at the discharge source to optimize the model. This enables density and dynamic viscosity to adaptively identify the dynamic evolution of physical properties caused by flash evaporation, solving the problem of fixed-parameter models failing due to physical property mismatch under flash evaporation conditions. This improves the adaptive identification capability of physical property parameters, ultimately outputting physically consistent three-dimensional velocity, pressure, density, and dynamic viscosity fields, thus improving overall robustness under complex operating conditions. Attached Figure Description
[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0011] Figure 1 A schematic diagram of the real-time reconstruction method of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision provided in an embodiment of this application; Figure 2 This is a schematic diagram of a physical information neural network model structure provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation
[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0013] During the injection phase of the safety relief valve, the fluid domain operates under extreme variable conditions, especially the flashing phenomenon caused by a sudden pressure drop, which leads to drastic dynamic changes in fluid density and dynamic viscosity. Traditional visual monitoring methods can only capture the liquid phase characteristics of the light-transmitting area, making it difficult to perceive the areas where visual observation data is missing due to dense atomization and the inherent evolution of physical parameters. Therefore, there is an urgent need for a three-dimensional physical field reconstruction method that can integrate fluid dynamics physical constraints, has the ability to fill in visually missing areas, and can perform inverse identification of physical properties.
[0014] To address this issue, this application provides a real-time reconstruction method for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural networks and multi-plane vision. This method solves the problems of missing visual observation data due to fluid viscosity, lack of physical constraints in traditional interpolation algorithms, and difficulty in capturing the dynamic evolution of physical properties under high-pressure discharge and flash evaporation conditions. This application embeds the fluid dynamics control equations into a physical information neural network, transforming the flow field reconstruction from pure data fitting to a fusion of physical constraints and observational data, ensuring the physical consistency, accuracy, and real-time performance of the reconstructed physical field under complex conditions.
[0015] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0016] In one exemplary embodiment, such as Figure 1 As shown, a real-time reconstruction method for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is described using a server as an example, and includes the following steps: Step S1: Real-time acquisition of multi-plane visual observation data and process parameters of the discharge source within the observation area to be reconstructed, and selection of multiple physical points within the observation area to be reconstructed; the multi-plane visual observation data includes droplet dynamic sequence images of multiple discrete observation planes within the observation area to be reconstructed, acquired through the multi-plane visual acquisition system. Step S2: Preprocess the multi-plane visual observation data to obtain the spatiotemporal coordinates of each visual observation point in the observation area to be reconstructed and the observed values of the droplet's motion velocity. Step S3: Input the spatiotemporal coordinates of each visual observation point and each physical coordinate point in the observation area to be reconstructed into the initial gas-liquid two-phase discharge physical field reconstruction model, and output the initial predicted values of the physical field variables of each visual observation point and each physical coordinate point in the observation area to be reconstructed. The initial gas-liquid two-phase discharge physical field reconstruction model is obtained by training a Physical Information Neural Network (PINNs) using a training set. The physical field variables include: three-dimensional velocity vector, pressure, density, and dynamic viscosity. Step S4: Based on the initial predicted values of the physical field variables at each visual observation point, the initial predicted values of the physical field variables at each physical alignment point, the observed values of the droplet velocity, and the process parameters at the discharge source, construct a composite loss function; the composite loss function includes: data observation loss, physical residual loss, initial condition loss, and boundary condition loss. Step S5: Optimize the network parameters of the initial gas-liquid two-phase discharge physical field reconstruction model with the composite loss function as the objective until the composite loss function converges, and obtain the target gas-liquid two-phase discharge physical field reconstruction model. Step S6: Input the spatiotemporal coordinates of the observation area to be reconstructed into the target gas-liquid two-phase discharge physical field reconstruction model, and output the predicted value of the three-dimensional physical field of the observation area to be reconstructed.
[0017] Specifically, the synergistic application of a multi-plane visual acquisition system and a physical information neural network model provides a new approach for real-time reconstruction of the physical field of gas-liquid two-phase discharge. A high-speed camera array, through a multi-plane layout, acquires dynamic sequence images of droplets from multiple discrete observation planes within the observation area to be reconstructed. A synchronization trigger ensures that the dynamic sequence images of the droplets are aligned with the process parameters acquired by pressure and temperature sensors according to a unified time reference, providing synchronized data for subsequent construction of data observation losses and determination of initial and boundary condition losses.
[0018] Building upon this foundation, this application achieves a leap from data fitting to physical inference by constructing a physical information neural network model. The physical information neural network model takes spatiotemporal coordinates as input and outputs the three-dimensional velocity field, pressure field, and dynamically evolving density and dynamic viscosity fields of the observation region to be reconstructed. Its core lies in the construction of a composite loss function, which includes not only the data observation loss based on visual observation data but also the physical residual loss corresponding to the mass and momentum conservation equations. Through automatic differentiation techniques, physical laws are transformed into constraints for model optimization, enabling the model to perform consistent inference based on the mass and momentum conservation equations even in regions lacking visual observation data, thus achieving occlusion compensation for the flow field structure.
[0019] As an optional implementation method, step S1 specifically includes: Step S11: Acquire dynamic sequence images of droplets on multiple discrete observation planes within the observation area to be reconstructed using multiple high-speed cameras in the multi-plane vision acquisition system. Step S12: The process parameters at the venting source are collected in real time by pressure and temperature sensors installed at the venting source. The process parameters include at least one of the following: pressure, temperature, component ratio of the material in the reactor, and opening degree of the safety venting valve. Step S13: Align the droplet dynamic sequence image with the process parameters using a synchronization trigger at a unified time reference; Step S14: Select multiple spatiotemporal coordinate points as physical points within the spatiotemporal domain of the observation area to be reconstructed.
[0020] Specifically, the multi-plane vision acquisition system adopts an array layout. By adjusting the object distance, focal length, and observation direction of multiple high-speed cameras, the observation planes are arranged in parallel, staggered, or fan-shaped distributions along the fluid discharge direction, forming multi-plane coverage of the discharge fluid spatial domain. Process parameters are used to determine the initial and boundary conditions of the physical information neural network and participate in the loss function calculation. To facilitate subsequent PINNs model training, the multi-plane vision acquisition system needs to establish a time synchronization triggering mechanism to ensure that high-speed image data and system pressure and temperature sensor data correspond to a unified time base, providing spatiotemporally correlated input data for subsequent data observation constraints and the construction of initial and boundary conditions.
[0021] As an optional implementation method, step S2 specifically includes: Step S21: Perform droplet identification and segmentation on the multi-frame dynamic sequence images of droplets to obtain the droplet region and the coordinates of the droplet region in the corresponding frame dynamic sequence image of each droplet in each discrete observation plane. Step S22: Based on the coordinates of the droplet region, perform temporal correlation on multiple frames of droplet dynamic sequence images to obtain the correspondence of the same droplet at different times and obtain the temporal trajectory of each droplet; Step S23: Based on the spatial calibration relationship between each discrete observation plane and the physical coordinate system of the observation area to be reconstructed, map the coordinates of the droplet region in the temporal trajectory of each droplet to the physical coordinate system to obtain the spatiotemporal coordinates of each visual observation point. Step S24: Calculate the velocity of the droplet at different times based on the spatiotemporal coordinates of the same droplet, and obtain the observed values of the droplet velocity at each visual observation point.
[0022] Specifically, in the image preprocessing stage, when preprocessing the multi-plane visual observation data, droplet identification and segmentation, temporal correlation and spatial coordinate mapping are performed on the multi-frame dynamic sequence images of droplets. The droplet regions and their coordinates obtained from each discrete observation plane are transformed into the spatiotemporal coordinates of each visual observation point and the observed values of droplet motion velocity in the physical coordinate system of the observation region to be reconstructed.
[0023] Specifically, a target detection algorithm is used to identify and segment droplets in a multi-frame dynamic sequence image of droplets, obtaining the droplet region and coordinates of each droplet in each discrete observation plane; based on the coordinates of the droplet region, a temporal correlation analysis is performed to obtain the temporal trajectory of each droplet, and finally the instantaneous velocity vector of the droplets on different two-dimensional planes (i.e., the observed value of the droplet's motion velocity) is determined.
[0024] As an optional implementation, in step S3, such as Figure 2 As shown, the physical information neural network includes an input layer, multiple fully connected hidden layers, and an output layer; The input layer is used to receive the spatiotemporal coordinates of each visual observation point or each physical coordinate point; A fully connected hidden layer is used to perform layer-by-layer nonlinear mapping of the spatiotemporal coordinates to obtain the high-dimensional spatiotemporal characteristics of the gas-liquid two-phase discharge flow field. The output layer is used to map high-dimensional spatiotemporal features and output the physical field variables.
[0025] Specifically, constructing a spatiotemporal coordinate system As input, with velocity field Pressure field A neural network that outputs physical information about fluid evolution parameters (density field ρ, dynamic viscosity field μ). The spatiotemporal coordinates of each visual observation point and the spatiotemporal coordinates of each physical point set within the spatiotemporal domain of the observation area to be reconstructed are used as network inputs. Through network mapping, predicted values of pressure, three-dimensional velocity vector, density, and dynamic viscosity at the corresponding spatiotemporal locations are obtained; where, These represent the three spatial coordinates in the physical coordinate system of the observation area after spatial calibration; t represents the corresponding time coordinate. They represent the speeds at... The directional components are established using continuous spatiotemporal coordinates. The mapping relationship to physical field variables.
[0026] The core innovation of the Physical Information Neural Network (PIN) lies in utilizing automatic differentiation technology to calculate the partial derivatives of physical field variables with respect to the input spatiotemporal coordinates. This directly embeds the physical equations into the model optimization process, eliminating the need for meshing complex geometric boundaries and avoiding convergence issues caused by poor mesh quality or discretization errors. Leveraging the continuous mapping characteristics of neural networks and the analytical derivative capabilities obtained through automatic differentiation, the model can achieve meshless solutions in the spatiotemporal continuous domain, significantly reducing preprocessing costs. Through the nonlinear transformation of the adaptive activation function in the hidden layer, it can accurately capture subtle local pressure fluctuations and rapid velocity gradient changes in the flow field during high-speed venting and transient flash evaporation, achieving physically consistent physical field reconstruction in data-sparse or visually obscured regions.
[0027] As an optional implementation, step S4 specifically includes: Step S41: Based on the deviation between the initial predicted value of the three-dimensional velocity vector of each visual observation point and the observed value of the droplet motion velocity at the corresponding visual observation point, construct the data observation loss. Step S42: Substitute the initial predicted values of the corresponding physical field variables of each physical collocation point into the mass conservation equation and the momentum conservation equation respectively to obtain the mass conservation residual and momentum conservation residual at each physical collocation point, and construct the physical residual loss based on the mass conservation residual and the momentum conservation residual. Step S43: Based on the process parameters of the discharge source, determine the initial conditions and boundary conditions of the observation area to be reconstructed, and construct the initial condition loss based on the initial conditions and the boundary condition loss based on the boundary conditions. Step S44: The data observation loss, the physical residual loss, the initial condition loss, and the boundary condition loss are weighted and summed according to preset weights to obtain the composite loss function.
[0028] As an optional implementation, in step S4, the expression for the composite loss function is: (1) (2) (3) in, It is a composite loss function; This is the physical residual loss; For data observation loss; Loss due to initial conditions; For boundary condition loss; , , and These are the weighting coefficients corresponding to each loss term; The number of physical points; Index for physical coordinate points; For the first The mass conservation residual of each physical collocation point; No. The momentum conservation residual of each physical collocation point; The square of the L2 norm; The number of visual observation points; For indexing visual observation points; For parameters The physical information neural network in the first The predicted value of the three-dimensional velocity vector output at each visual observation point; For the first Projection operator corresponding to each visual observation point; For the first The observed values of droplet velocity at each visual observation point.
[0029] Specifically, physical collocation points are selected within the spatiotemporal domain corresponding to the observation area to be reconstructed. Automatic differentiation is used to obtain the partial derivatives of the physical field variables output by the initial gas-liquid two-phase discharge physical field reconstruction model with respect to the spatiotemporal coordinates. The physical field variables and their partial derivatives are substituted into the mass conservation equation and the momentum conservation equation to obtain the mass conservation residuals and momentum conservation residuals at each physical collocation point. The mass conservation residual is: (4) in, For quality conservation residuals; Density; It is a three-dimensional velocity vector.
[0030] The momentum conservation residual is: ( 5) in, For momentum conservation residuals, For tensor product operators; Let μ be the viscous stress tensor determined by the dynamic viscosity μ and the velocity gradient. This is the gravitational acceleration vector.
[0031] The physical residual loss is constructed based on the mass conservation residual and momentum conservation residual at each physical collocation point, as shown in Equation (2).
[0032] At the visual observation point, based on the spatial calibration relationship of the corresponding observation plane, the three-dimensional velocity vector output by the initial gas-liquid two-phase discharge physical field reconstruction model is converted into the predicted velocity component in the observation plane. The deviation between the predicted velocity component and the observed value of the droplet motion velocity at the corresponding visual observation point is calculated, and the data observation loss is constructed as shown in Equation (3).
[0033] The physical residual loss, data observation loss, initial condition loss, and boundary condition loss are weighted and summed according to preset weights to obtain the composite loss function used for iterative optimization of the initial gas-liquid two-phase venting physical field reconstruction model, as shown in Equation (1).
[0034] The composite loss function integrates data observation constraints with fluid dynamics control equation constraints to establish a multi-dimensional physical constraint mechanism, which includes the mass conservation equation and momentum conservation equation describing the basic laws of fluid motion, as well as the initial conditions and boundary conditions determined by the process parameters at the discharge source.
[0035] When the flow field is in a high-concentration spray area where the monitoring line of sight is obstructed or a sampling blind area where the data samples are extremely sparse, no corresponding data observation loss is constructed for the location where effective visual data cannot be obtained. Instead, physical residual constraints are used to constrain the network output of the region by combining the initial conditions, boundary conditions and the adjacent effective visual observation data. This achieves physical consistency completion and occlusion compensation for the missing flow field structure, ensuring that the reconstructed pressure field, velocity field, density field and dynamic viscosity field maintain physical consistency within the observation area.
[0036] As an optional implementation, step S5 specifically includes: Step S51: The Adam optimization algorithm is used to globally optimize the network parameters of the initial gas-liquid two-phase discharge physical field reconstruction model. Step S52: When the relative change of the composite loss function in the continuous iteration interval is less than the preset switching threshold, the Adam optimization algorithm is switched to the L-BFGS optimization algorithm to perform second-order optimization on the network parameters of the initial gas-liquid two-phase discharge physical field reconstruction model until the composite loss function converges, and the target gas-liquid two-phase discharge physical field reconstruction model is obtained.
[0037] Specifically, the network parameter optimization of the initial gas-liquid two-phase venting physical field reconstruction model adopts the following strategy: using the composite loss function as the network optimization objective, and taking the dynamically evolving density and dynamic viscosity during flash evaporation as the physical property variables updated with network parameter optimization, the Adam optimization algorithm is used for global search to quickly reduce the composite loss function value; after the composite loss function value enters a stable period (i.e., the relative change within the continuous iteration interval is less than the preset switching threshold), the L-BFGS optimization algorithm is switched to perform second-order fine optimization to improve the convergence accuracy of the flow field reconstruction until the composite loss function converges. During the optimization process, the inverse identification of fluid density and dynamic viscosity and model calibration are achieved by minimizing the composite loss function, thus obtaining the target gas-liquid two-phase venting physical field reconstruction model.
[0038] By employing a hybrid optimization strategy combining the Adam and L-BFGS algorithms, an optimization process is constructed from global coarse-grained optimization to local fine-grained optimization through phased logical switching. In the initial optimization phase, the adaptive learning rate characteristic and first-order momentum estimation capability of the Adam algorithm are utilized to perform gradient optimization within the high-dimensional nonlinear loss space, rapidly reducing the composite loss function. When the relative change of the composite loss function within a preset continuous iteration interval is less than a preset switching threshold, the composite loss function is determined to have entered a stable phase, and the algorithm automatically switches to the L-BFGS algorithm based on the quasi-Newton method. Leveraging its approximate representation of curvature information, this algorithm further reduces physical residual loss and data observation loss. During this optimization process, density and dynamic viscosity are used as variables to be inverted. Their spatiotemporal distribution within the observation region to be reconstructed is obtained through iterative optimization of network parameters, enabling inverse identification of fluid property changes during flash evaporation and model calibration, thus improving the model's convergence stability under complex conditions.
[0039] As an optional implementation, step S6 specifically includes: Using the target gas-liquid two-phase discharge physical field reconstruction model obtained in step S5, physical field inference is performed on the continuous spatiotemporal coordinates within the observation area to be reconstructed to obtain the three-dimensional velocity field, pressure field, density field, and dynamic viscosity field. For areas with valid visual observation data, the model output is determined jointly by visual observation data constraints and physical residual constraints. For areas where visual observation data is missing due to droplet concentration, overlap, or occlusion, the physical field of the area is uniformly completed by the physical residual constraints corresponding to the mass conservation equation and momentum conservation equation, combined with the initial conditions, boundary conditions, and adjacent valid visual observation data, to obtain the continuous three-dimensional physical field within the observation area to be reconstructed.
[0040] As an optional implementation, the predicted values of the three-dimensional physical field include: predicted values of the three-dimensional velocity field, predicted values of the pressure field, predicted values of the density field, and predicted values of the dynamic viscosity field; the method further includes: Step S7: Based on the predicted values of the three-dimensional velocity field and the three-dimensional density field of the observation area to be reconstructed, determine the real-time mass release rate of the gas-liquid two-phase release process.
[0041] As an optional implementation, step S7 specifically includes: Step S71: Select an integral section spanning the discharge flow region within the observation area to be reconstructed, and extract the predicted values of the three-dimensional velocity field and density field on the integral section. Step S72: Multiply the predicted value of the three-dimensional velocity field along the normal direction of the integral section with the predicted value of the three-dimensional density field to obtain the mass flux on the integral section. Step S73: Perform area integral on the mass flux on the integral section to obtain the real-time mass discharge rate of the gas-liquid two-phase mixed discharge flow.
[0042] Specifically, the formula for calculating the real-time quality leakage rate is as follows: (6) in, for Real-time quality discharge rate at any given moment; This is the integral section spanning the venting flow region; Let be the unit normal vector of the integral section; Position on the integration section At any moment The corresponding predicted value of the three-dimensional density field, Position on the integration section At any moment The corresponding predicted value of the three-dimensional velocity field.
[0043] In this application, a target gas-liquid two-phase discharge physical field reconstruction model is used to perform analytical reasoning in the continuous spatiotemporal domain to reconstruct the continuous three-dimensional velocity field, pressure field, density field, and dynamic viscosity field within the observation area to be reconstructed. For areas lacking effective visual observation data due to droplet concentration, droplet overlap, or line-of-sight obstruction, the physical field within the area is continuously reconstructed using the mass conservation equation and momentum conservation equation corresponding to the constructed physical residual loss, combined with initial conditions, boundary conditions, and effective visual observation data from neighboring areas. By establishing a continuous physical field function based on spatiotemporal coordinates, the mass flux formed by the velocity components of the predicted values of the three-dimensional density field and the three-dimensional velocity field along the normal direction of the selected integral section is calculated using surface integral. This enables online calculation of the real-time mass discharge rate of the gas-liquid two-phase mixed discharge flow, reducing the loss of discharge monitoring information caused by local invisibility in traditional visual methods, and providing data support for real-time monitoring and subsequent safety control of the discharge process.
[0044] As an optional implementation, the method further includes: Step S8: The real-time mass discharge rate is transmitted to the PLC controller, which then dynamically adjusts the opening of the safety relief valve based on the real-time mass discharge rate.
[0045] To verify the technical effectiveness of the real-time reconstruction method for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural networks and multi-plane vision provided in this application, the following description is provided in conjunction with specific application scenarios. The following embodiments are for illustrative purposes only and do not constitute a limitation on the scope of protection of this application.
[0046] Example 1 This embodiment is applied to the monitoring of the emergency venting system of a high-pressure reactor in a chemical plant. The initial venting pressure is 5 MPa and the temperature is 150°C. At the time of venting, the multi-plane vision acquisition system captures the initial morphology of the jet flow. Due to the extremely high atomization concentration near the nozzle, traditional PIV technology cannot obtain velocity information in the central region. This embodiment uses a physical information neural network model, utilizing visual observation data from the peripheral discrete observation planes, combined with the physical residual constraints corresponding to the mass conservation equation and momentum conservation equation, to successfully calculate the pressure gradient and three-dimensional velocity field in the region lacking visual observation data. The hybrid optimization algorithm switches to the L-BFGS optimization algorithm after 500 Adam iterations, minimizing the physical residual loss. The accuracy was measured at an order of magnitude, revealing a density decrease of approximately 40% in the flash region. The reconstructed real-time mass release rate was integrated over the release time to obtain the cumulative release mass. The error between the obtained cumulative release mass and the measured value in the subsequent collection tank was only 2.8%, significantly better than the 15% error of traditional interpolation methods.
[0047] To verify the applicability of this application under ultra-high pressure and rapid property change venting conditions, Example 2 is further provided. The main difference between Example 2 and Example 1 is the different venting conditions. The reconstruction method is the same as steps S1 to S8 above, so the same parts will not be repeated.
[0048] Example 2 Under ultra-high pressure (10MPa) venting conditions, fluid properties change extremely rapidly. In this embodiment, step S4 introduces the physical residual loss corresponding to the mass conservation equation and momentum conservation equation. Combining the initial conditions, boundary conditions, and multi-plane visual observation data at the venting source, a composite loss function is constructed and the initial gas-liquid two-phase venting physical field reconstruction model is optimized. Experimental results show that, even with extremely sparse visual observation data (only 20% visibility), the target gas-liquid two-phase venting physical field reconstruction model utilizes physical residual constraints and effective visual observation data to continuously reconstruct the three-dimensional velocity field, pressure field, density field, and dynamic viscosity field in the region lacking visual observation data. Based on the reconstructed three-dimensional density field and three-dimensional velocity field, the real-time mass venting rate is calculated, providing real-time data support for the PLC controller to dynamically adjust the opening of the safety venting valve.
[0049] Beneficial effects: This application utilizes a reverse problem-solving mechanism to treat density and dynamic viscosity as physical property field variables updated with network parameter optimization. It automatically identifies and corrects the spatiotemporal evolution parameters of fluid density and dynamic viscosity under flash evaporation conditions in real time, achieving adaptive identification of physical property parameters. By leveraging the global constraint capability of the physical residual constraints corresponding to the mass conservation equation and momentum conservation equation, it consistently completes the regions where visual observation data is missing due to droplet concentration, overlap, or occlusion, thereby achieving occlusion compensation of the flow field structure.
[0050] 1) Improve physical consistency: By introducing the physical residual loss corresponding to the mass conservation equation and momentum conservation equation into the composite loss function, the reconstructed three-dimensional velocity field, pressure field, density field and dynamic viscosity field satisfy the basic laws of fluid mechanics, avoiding the output of non-physical error solutions by the pure data-driven model.
[0051] 2) Improve occlusion compensation capability: By selecting physical points within the observation area to be reconstructed, the physical residual constraints cover the areas where visual observation data is missing. Thus, when droplet concentration, overlap, or line-of-sight occlusion causes visual observation data to be missing, the physical field of the missing area can still be consistently completed.
[0052] 3) Improve the adaptive identification capability of physical property parameters: By outputting density and dynamic viscosity as physical field variables, and constructing a composite loss function based on the visual observation value, physical collocation point prediction value and discharge source process parameters under the current working conditions, density and dynamic viscosity can be adaptively identified as physical property dynamics evolve due to flash evaporation.
[0053] 4) Improve robustness under complex operating conditions: By combining multi-plane visual observation data with physical information neural networks and using composite loss functions to optimize the model, the model can adapt to complex discharge conditions such as high-pressure flashing and drastic changes in physical properties, thereby improving the stability and reliability of the reconstruction process.
[0054] 5) Improve reconstruction accuracy and parameter identification efficiency: By constraining both data observation loss and physical residual loss, and combining the Adam optimization algorithm and the L-BFGS optimization algorithm in a hybrid optimization strategy, the flow field reconstruction accuracy and physical property parameter identification efficiency under high-pressure flash evaporation conditions are improved.
[0055] 6) Reduce hardware dependence and data acquisition costs: No need for massive labeled samples, physical field reconstruction can be achieved under conditions where visual observation data is sparse or partially missing, reducing dependence on high-performance hardware environment and intensive data acquisition.
[0056] 7) Expanding industrial application prospects: It can be applied to industrial scenarios such as petrochemical safety monitoring and high-pressure reactor venting assessment, realizing continuous and stable acquisition of physical information of the entire gas-liquid two-phase flow field, with significant safety benefits, economic benefits and broad industrial application prospects.
[0057] Based on the same inventive concept, this application also provides a system for real-time reconstruction of the physical field of gas-liquid two-phase discharge, which is based on the fusion of physical information neural networks and multi-plane vision, to implement the aforementioned method for real-time reconstruction of the physical field of gas-liquid two-phase discharge. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the real-time reconstruction system for gas-liquid two-phase discharge physical field based on the fusion of physical information neural networks and multi-plane vision provided below can be found in the limitations of the method for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural networks and multi-plane vision described above, and will not be repeated here.
[0058] In one exemplary embodiment, a real-time reconstruction system for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural networks and multi-plane vision is provided, comprising: The data acquisition and physical point selection unit is used to acquire multi-plane visual observation data and process parameters of the discharge source in the observation area to be reconstructed in real time, and select multiple physical points in the observation area to be reconstructed; the multi-plane visual observation data includes droplet dynamic sequence images of multiple discrete observation planes in the observation area to be reconstructed, acquired by the multi-plane visual acquisition system. The preprocessing unit is used to preprocess the multi-plane visual observation data to obtain the spatiotemporal coordinates of each visual observation point in the observation area to be reconstructed and the observed values of the droplet motion velocity. The initial reconstruction unit is used to input the spatiotemporal coordinates of each visual observation point and each physical coordinate point in the observation area to be reconstructed into the initial gas-liquid two-phase discharge physical field reconstruction model, and output the initial predicted values of the physical field variables of each visual observation point and each physical coordinate point in the observation area to be reconstructed. The initial gas-liquid two-phase discharge physical field reconstruction model is obtained by training a physical information neural network using a training set. The physical field variables include: three-dimensional velocity vector, pressure, density, and dynamic viscosity. The loss function construction unit is used to construct a composite loss function based on the initial predicted values of the physical field variables at each visual observation point, the initial predicted values of the physical field variables at each physical alignment point, the observed values of the droplet velocity, and the process parameters at the discharge source. The composite loss function includes: data observation loss, physical residual loss, initial condition loss, and boundary condition loss. The model optimization unit is used to optimize the network parameters of the initial gas-liquid two-phase discharge physical field reconstruction model with the composite loss function as the objective until the composite loss function converges, thus obtaining the target gas-liquid two-phase discharge physical field reconstruction model. The target reconstruction unit is used to input the spatiotemporal coordinates of the observation area to be reconstructed into the target gas-liquid two-phase discharge physical field reconstruction model, and output the predicted value of the three-dimensional physical field of the observation area to be reconstructed.
[0059] In one exemplary embodiment, a computer device is provided, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a real-time reconstruction method for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural networks and multi-plane vision.
[0060] In one exemplary embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements a method for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural networks and multi-plane vision.
[0061] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements a real-time reconstruction method for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural networks and multi-plane vision.
[0062] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 3 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a real-time reconstruction method for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural networks and multi-plane vision.
[0063] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0064] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.
[0065] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0066] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0067] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0068] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods, systems, and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision, characterized in that, The method includes: The system acquires multi-plane visual observation data and process parameters of the discharge source within the observation area to be reconstructed in real time, and selects multiple physical points within the observation area to be reconstructed; the multi-plane visual observation data includes droplet dynamic sequence images of multiple discrete observation planes within the observation area to be reconstructed, acquired through the multi-plane visual acquisition system. Preprocessing of multi-plane visual observation data yields the spatiotemporal coordinates of each visual observation point within the observation area to be reconstructed, as well as the observed values of droplet velocity. The spatiotemporal coordinates of each visual observation point and each physical coordinate point in the observation area to be reconstructed are input into the initial gas-liquid two-phase discharge physical field reconstruction model, which outputs the initial predicted values of the physical field variables of each visual observation point and each physical coordinate point in the observation area to be reconstructed. The initial gas-liquid two-phase discharge physical field reconstruction model is obtained by training a physical information neural network using a training set. The physical field variables include: three-dimensional velocity vector, pressure, density, and dynamic viscosity. Based on the initial predicted values of the physical field variables at each visual observation point, the initial predicted values of the physical field variables at each physical alignment point, the observed values of the droplet velocity, and the process parameters at the discharge source, a composite loss function is constructed. The composite loss function includes: data observation loss, physical residual loss, initial condition loss, and boundary condition loss. The network parameters of the initial gas-liquid two-phase venting physical field reconstruction model are optimized with the composite loss function as the objective until the composite loss function converges, thus obtaining the target gas-liquid two-phase venting physical field reconstruction model. The spatiotemporal coordinates of the observation area to be reconstructed are input into the target gas-liquid two-phase discharge physical field reconstruction model, and the predicted value of the three-dimensional physical field of the observation area to be reconstructed is output.
2. The method for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision as described in claim 1, characterized in that, Real-time acquisition of multi-plane visual observation data and process parameters of the discharge source within the observation area to be reconstructed, and selection of multiple physical points within the observation area to be reconstructed, specifically including: The system uses multiple high-speed cameras in a multi-plane vision acquisition system to acquire dynamic sequence images of droplets on multiple discrete observation planes within the observation area to be reconstructed. The process parameters at the venting source are collected in real time by pressure and temperature sensors installed at the venting source. The process parameters include at least one of the following: pressure, temperature, component ratio of the material in the reactor, and opening degree of the safety venting valve. The droplet dynamic sequence image is aligned with the process parameters using a synchronous trigger at a unified time reference. Multiple spatiotemporal coordinate points are selected as physical points within the spatiotemporal domain of the observation area to be reconstructed.
3. The method for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision as described in claim 1, characterized in that, Preprocessing of multi-plane visual observation data yields the spatiotemporal coordinates of each visual observation point within the observation area to be reconstructed, as well as the observed values of droplet velocity. Specifically, this includes: Droplet identification and segmentation are performed on multiple frames of dynamic sequence images of droplets to obtain the droplet region and coordinates of each droplet in the corresponding frame of dynamic sequence image of droplets in each discrete observation plane; Based on the coordinates of the droplet region, the dynamic sequence images of multiple droplets are temporally correlated to obtain the correspondence of the same droplet at different times and obtain the temporal trajectory of each droplet. Based on the spatial calibration relationship between each discrete observation plane and the physical coordinate system of the observation area to be reconstructed, the coordinates of the droplet region in the temporal trajectory of each droplet are mapped to the physical coordinate system to obtain the spatiotemporal coordinates of each visual observation point. Based on the spatiotemporal coordinates of the same droplet at different times, the velocity of the corresponding droplet is calculated, and the observed values of the droplet velocity at each visual observation point are obtained.
4. The method for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision as described in claim 1, characterized in that, A physical information neural network consists of an input layer, multiple fully connected hidden layers, and an output layer. The input layer is used to receive the spatiotemporal coordinates of each visual observation point or each physical coordinate point; A fully connected hidden layer is used to perform layer-by-layer nonlinear mapping of the spatiotemporal coordinates to obtain the high-dimensional spatiotemporal characteristics of the gas-liquid two-phase discharge flow field. The output layer is used to map high-dimensional spatiotemporal features and output the physical field variables.
5. The method for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision as described in claim 1, characterized in that, Based on the initial predicted values of the physical field variables at each visual observation point, the initial predicted values of the physical field variables at each physical alignment point, the observed values of the droplet velocity, and the process parameters at the discharge source, a composite loss function is constructed, specifically including: Based on the deviation between the initial predicted value of the three-dimensional velocity vector at each visual observation point and the observed value of the droplet motion velocity at the corresponding visual observation point, a data observation loss is constructed. Substitute the initial predicted values of the corresponding physical field variables at each physical collocation point into the mass conservation equation and the momentum conservation equation respectively to obtain the mass conservation residual and momentum conservation residual at each physical collocation point, and construct the physical residual loss based on the mass conservation residual and the momentum conservation residual. Based on the process parameters at the discharge source, the initial conditions and boundary conditions of the observation area to be reconstructed are determined, and the initial condition loss is constructed based on the initial conditions, and the boundary condition loss is constructed based on the boundary conditions. The composite loss function is obtained by weighting and summing the data observation loss, the physical residual loss, the initial condition loss, and the boundary condition loss according to preset weights.
6. The real-time reconstruction method for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision as described in claim 5, characterized in that, The expression for the composite loss function is: ; ; ; in, It is a composite loss function; This is the physical residual loss; For data observation loss; Loss due to initial conditions; For boundary condition loss; , , and These are the weighting coefficients corresponding to each loss term; The number of physical points; Index for physical coordinate points; For the first The mass conservation residual of each physical collocation point; No. The momentum conservation residual of each physical collocation point; The square of the L2 norm; The number of visual observation points; For indexing visual observation points; For parameters The physical information neural network in the first The predicted value of the three-dimensional velocity vector output at each visual observation point; For the first Projection operator corresponding to each visual observation point; For the first The observed values of droplet velocity at each visual observation point.
7. The method for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision according to claim 1, characterized in that, Using the composite loss function as the objective, the network parameters of the initial gas-liquid two-phase venting physical field reconstruction model are optimized until the composite loss function converges, resulting in the target gas-liquid two-phase venting physical field reconstruction model, which specifically includes: The Adam optimization algorithm is used to globally optimize the network parameters of the initial gas-liquid two-phase discharge physical field reconstruction model; When the relative change of the composite loss function within the continuous iteration interval is less than the preset switching threshold, the Adam optimization algorithm is switched to the L-BFGS optimization algorithm to perform second-order optimization on the network parameters of the initial gas-liquid two-phase venting physical field reconstruction model until the composite loss function converges, thus obtaining the target gas-liquid two-phase venting physical field reconstruction model.
8. The method for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision according to claim 1, characterized in that, The predicted values for the three-dimensional physics field include: the predicted values for the three-dimensional velocity field, pressure field, density field, and dynamic viscosity field. The method further includes: Based on the predicted values of the three-dimensional velocity field and the three-dimensional density field of the observation area to be reconstructed, the real-time mass release rate of the gas-liquid two-phase release process is determined.
9. The method for real-time reconstruction of the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision according to claim 8, characterized in that, Based on the predicted values of the three-dimensional velocity field and the three-dimensional density field of the observation area to be reconstructed, the real-time mass release rate of the gas-liquid two-phase release process is determined, specifically including: In the observation area to be reconstructed, an integral section spanning the discharge flow region is selected, and the predicted values of the three-dimensional velocity field and density field on the integral section are extracted. The predicted value of the three-dimensional velocity field along the normal direction of the integral section is multiplied by the predicted value of the three-dimensional density field to obtain the mass flux on the integral section. The mass flux on the integral section is integraled over the integral section to obtain the real-time mass discharge rate of the gas-liquid two-phase mixed discharge flow.
10. A real-time reconstruction system for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision, characterized in that, The real-time reconstruction system for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision is used to implement the real-time reconstruction method for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision as described in any one of claims 1-9. The real-time reconstruction system for the physical field of gas-liquid two-phase discharge based on the fusion of physical information neural network and multi-plane vision includes: The data acquisition and physical point selection unit is used to acquire multi-plane visual observation data and process parameters of the discharge source in the observation area to be reconstructed in real time, and select multiple physical points in the observation area to be reconstructed; the multi-plane visual observation data includes droplet dynamic sequence images of multiple discrete observation planes in the observation area to be reconstructed, acquired by the multi-plane visual acquisition system. The preprocessing unit is used to preprocess the multi-plane visual observation data to obtain the spatiotemporal coordinates of each visual observation point in the observation area to be reconstructed and the observed values of the droplet motion velocity. The initial reconstruction unit is used to input the spatiotemporal coordinates of each visual observation point and each physical coordinate point in the observation area to be reconstructed into the initial gas-liquid two-phase discharge physical field reconstruction model, and output the initial predicted values of the physical field variables of each visual observation point and each physical coordinate point in the observation area to be reconstructed. The initial gas-liquid two-phase discharge physical field reconstruction model is obtained by training a physical information neural network using a training set. The physical field variables include: three-dimensional velocity vector, pressure, density, and dynamic viscosity. The loss function construction unit is used to construct a composite loss function based on the initial predicted values of the physical field variables at each visual observation point, the initial predicted values of the physical field variables at each physical alignment point, the observed values of the droplet velocity, and the process parameters at the discharge source. The composite loss function includes: data observation loss, physical residual loss, initial condition loss, and boundary condition loss. The model optimization unit is used to optimize the network parameters of the initial gas-liquid two-phase discharge physical field reconstruction model with the composite loss function as the objective until the composite loss function converges, thus obtaining the target gas-liquid two-phase discharge physical field reconstruction model. The target reconstruction unit is used to input the spatiotemporal coordinates of the observation area to be reconstructed into the target gas-liquid two-phase discharge physical field reconstruction model, and output the predicted value of the three-dimensional physical field of the observation area to be reconstructed.
