A method for reconstructing combustion state in a combustion engine control system

By constructing a physical information neural network, the high gradient region of the combustion chamber of the combustion engine is identified and reconstructed, solving the problem of multi-physics data distortion in the combustion control system, realizing high-precision combustion state reconstruction and control, and improving the reliability and control accuracy of the system.

CN122528685BActive Publication Date: 2026-11-10ZHEJIANG SCI-TECH UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610995025.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-07-06
Publication Date
2026-11-10
Estimated Expiration
2046-07-06

AI Technical Summary

Technical Problem

Existing technologies struggle to accurately reconstruct multiphysics data in the combustion chamber of combustion engines, especially in high-gradient regions such as the flame front, where significant distortion occurs, resulting in insufficient accuracy and reliability of the combustion control system.

Method used

A physical information neural network is constructed. By identifying high gradient regions such as flame fronts, high-fidelity, reference flame fronts, and distortion regions are divided. A total loss function is constructed that includes data loss, physical equation loss, and distortion correction loss. The network is then reconstructed using sparse observation data.

Benefits of technology

It achieves high-precision data reconstruction of the combustion control system, improves the accuracy of mixture ratio control, combustion stability control and heat load regulation, supports the full life cycle development and application of the combustion control system, and has good generalization ability and engineering application value.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122528685B_ABST
    Figure CN122528685B_ABST
Patent Text Reader

Abstract

This invention relates to the field of combustion engine control technology and discloses a method for reconstructing the combustion state in a combustion engine control system. The method involves acquiring distorted multiphysics field data samples from the combustion chamber, then dividing the combustion chamber space into a high-fidelity region, a reference flame front region, and a distortion region. A physical information neural network is constructed, incorporating data loss, physical equation loss, and distortion correction loss. Prior weights are assigned to each loss function based on the region to which the grid nodes belong. The network is then trained offline using samples and prior weights to obtain an optimized network. Real-time collected sparse observation data is input into the optimized network, and the reconstructed physical quantities are output. This invention provides high-fidelity, real-time, full-field multiphysics field reconstruction of the combustion state in combustion engine control systems, offering parameter inputs for closed-loop control and state monitoring.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of combustion engine control technology, and particularly relates to a method for reconstructing the combustion state in a combustion engine control system. Background Technology

[0002] The operation of an engine combustion chamber involves extremely complex physicochemical changes, and its combustion efficiency, stability, and thermal protection performance directly determine the engine's reliability and reusability. Achieving precise regulation and control of the engine combustion process (such as propellant mixture control, active suppression of combustion instability, and thermal protection management) highly depends on the accurate acquisition of key physical fields within the combustion chamber (such as temperature, component concentration, and velocity fields). Therefore, high-precision multiphysics data is a crucial foundation for realizing high-performance combustion control systems. However, in key "interface" regions such as the flame front, fuel / oxidizer shear mixing layer, and droplet evaporation interface, physical quantities such as component concentration, temperature, and pressure undergo drastic gradient changes within a small spatial scale. The low precision in these regions restricts the development of high-precision combustion control technology.

[0003] Currently, numerical simulation based on Computational Fluid Dynamics (CFD) is the primary means of obtaining multiphysics data for combustion chambers. However, limited by computational resources and model assumptions, traditional CFD methods have significant shortcomings in handling the aforementioned interfaces: to balance engineering timeliness, CFD simulations often employ Reynolds-Averaged Navier-Stokes (RANS) turbulence models and simplified overall chemical reaction mechanisms. This simplification is acceptable in stable multiphysics fields far from the reaction zone, but in regions with strong chemical rigidity, such as the flame front, it cannot accurately capture turbulence-chemical interaction (TCI), leading to significant deviations between the calculated and actual values ​​of local physical quantities, forming "distorted" data areas. The characteristic scales of combustion reactions (such as laminar flame thickness) are typically on the sub-millimeter scale, while the macroscopic scale of engineering-grade combustion chambers can reach the meter scale. Due to limitations in the number of meshes, CFD meshes often struggle to resolve the fine structure of flames, leading to numerical dissipation or spurious numerical fluctuations (Gibbs phenomenon) at the front. Parameters required for simulation, such as inlet spray distribution and initial turbulence intensity, are typically based on idealized assumptions. Initial errors in these boundary conditions accumulate and amplify downstream in the multiphysics field (especially in the flame stability region, which is sensitive to initial conditions). Besides numerical simulation, experiments are another way to obtain data. However, in extreme environments of high temperature, high pressure, and strong radiation, the optical measurement window is limited. Existing measurement techniques (such as planar laser-induced fluorescence (PLIF) and particle image velocimetry (PIV) can only provide sparse, discontinuous point or two-dimensional slice data, failing to reconstruct a complete three-dimensional transient multiphysics field.

[0004] In recent years, data-driven methods based on deep learning (such as convolutional neural networks (CNN) and recurrent neural networks (RNN) have been used for multiphysics prediction. However, pure data-driven models have problems such as high dependence on high-fidelity training data, poor generalization ability, and lack of physical interpretability. They are prone to producing prediction results that violate physical laws in regions with strong physical constraints such as mass conservation or energy conservation.

[0005] Physics-Informed Neural Networks (PINNs) integrate physical laws with observational data by introducing the residuals of partial differential equations (PDEs) as regularization terms into the loss function. This allows for solving inverse problems and enhancing generalization capabilities with small sample sizes. However, current PINN techniques primarily address idealized problems and have not yet effectively solved the data reconstruction and completion problems in multi-physics, multiphase flow, and high-dimensional, strongly nonlinear systems such as liquid engine combustion chambers. In particular, there is a lack of efficient methods for real-time monitoring and control applications of combustion processes that can automatically focus on key distortion regions while ensuring physical consistency. Current technologies still fail to address the following key issues:

[0006] (1) Lack of reconstruction mechanism for local distortion: The existing PINN has difficulty in automatically identifying and focusing on the distortion region (such as flame front) in CFD simulation, resulting in the waste of computational resources in the flat region of the physical field, and the inability to effectively correct the numerical error in the high gradient region.

[0007] (2). Difficulty in weighting loss function: In the multiphysics field of combustion involving multi-component reactions, the residual magnitudes of each physical equation (momentum, energy, component transport) differ greatly, and the chemical reaction source term has strong rigidity, making it difficult for conventional fixed weight strategies to converge.

[0008] (3) Lack of ability to handle multiphase discontinuities: Existing continuity assumptions are difficult to apply directly to regions with phase interface discontinuities, such as droplet evaporation, resulting in a significant decrease in the training accuracy of the model in multiphase and multiphysics fields.

[0009] Therefore, there is an urgent need to develop a method that can specifically reconstruct data distortion in key combustion regions and reconstruct a high-fidelity physical field across the entire field using sparse observation data. Summary of the Invention

[0010] The purpose of this invention is to provide a method for reconstructing the combustion state in a combustion engine control system, which achieves high-precision reconstruction of multi-physics field data in the combustion chamber by constructing a physical information neural network.

[0011] To achieve the above objectives, the present invention provides a combustion state reconfiguration method for a combustion engine control system, comprising the following steps:

[0012] S1. Obtain distorted multiphysics data samples of the combustion chamber. The distorted multiphysics data samples include the spatial coordinates and physical quantities of the grid nodes. The physical quantities include at least the mass fraction of fuel components.

[0013] S2. Determine the center position of the flame front based on the spatial gradient of the fuel component mass fraction, determine the thickness of the distorted flame based on the variation of the fuel component mass fraction, and then divide the spatial region of the combustion chamber into the reference flame front region, the distorted region and the high-fidelity region.

[0014] S3. Construct a physical information neural network and a total loss function that includes data loss, physical equation loss, and distortion correction loss; assign prior weights to the loss of each grid node according to whether it belongs to the high-fidelity region, the reference flame front region, or the distortion region.

[0015] S4. Use distorted multiphysics data samples and prior weights to train the physical information neural network offline to obtain the optimized physical information neural network;

[0016] S5. Real-time acquisition of sparse observation data from the combustion chamber, input into the physical information neural optimization network, and output of reconstructed physical quantities.

[0017] As an improvement to the combustion state reconstruction method for a combustion engine control system of the present invention:

[0018] The method for acquiring distorted multiphysics data samples in step S1 includes:

[0019] (1). If the original data obtained is obtained by numerical simulation of the target combustion chamber operating conditions through CFD simulation, then the discrete multiphysics data at each grid node is extracted and then normalized.

[0020] (2). If the original data is obtained from sensor measurements, sparse data expansion is first performed using a temporary pre-trained physical information neural network that includes data loss and physical equation loss, and then normalization is performed.

[0021] As a further improvement to the combustion state reconstruction method for a combustion engine control system of the present invention:

[0022] The process of dividing the reference flame front region, distortion region, and high-fidelity region in step S2 includes:

[0023] (1) Calculate the first-order spatial gradient of fuel component mass fraction at each grid node. , Indicates the position of the mesh node along the flame normal;

[0024] (2). First-order spatial gradient The location of the maximum value is determined as the center of the flame front;

[0025] (3). The first-order spatial gradient Normalization is performed to obtain the normalized gradient field. The set of spatial locations in the normalized gradient field that are greater than the preset gradient threshold is taken as the flame front region.

[0026] (4). The region centered on the center of the flame front and with a thickness equal to the reference flame thickness is taken as the reference flame front region; the region corresponding to the distorted flame thickness is merged with the flame front region, and the high-fidelity region is the region outside the merged region. The part of the merged region other than the reference flame front region is taken as the distorted region.

[0027] As a further improvement to the combustion state reconstruction method for a combustion engine control system of the present invention:

[0028] The thickness of the distorted flame satisfy:

[0029]

[0030] in, This represents the 10% mark within the range of variations in the mass fraction of fuel components. This represents the 90% position within the range of changes in the mass fraction of fuel components.

[0031] The reference flame thickness satisfies:

[0032]

[0033] in, Indicates the reference flame thickness. For reference, the characteristic gradient of the flame front region:

[0034] ;

[0035] in, To reference the fuel component chemical reaction source term at the center of the flame front region, , For reference, the density and velocity at the center of the flame front region;

[0036] and Fuel component mass fractions at the 10% and 90% positions, respectively:

[0037] ,

[0038] ;

[0039] in, The mass fraction of fuel components in the unburned zone. This represents the mass fraction of fuel components at the combustion chamber outlet.

[0040] As a further improvement to the combustion state reconstruction method for a combustion engine control system of the present invention:

[0041] The data loss is:

[0042]

[0043] in, This represents the total number of grid points in the high-fidelity region, the reference flame front region, and the distortion region within three adjacent grid areas. ns represents the number of computed physical quantities output. Indicates the coordinates of the node The i-th calculated physical quantity output at point i. Represents the node coordinates in a distorted multiphysics data sample. The i-th physical quantity at point i, Indicates the coordinates of the node The data loss weight for the i-th physical quantity at point i:

[0044]

[0045] in, This represents the normalized gradient field.

[0046] As a further improvement to the combustion state reconstruction method for a combustion engine control system of the present invention:

[0047] The loss of the physical equation is:

[0048]

[0049] in, This represents the total number of grid points in the distorted region and the high-fidelity region. Representing node coordinates The residuals of the component conservation equation at the location, Representing node coordinates The physical equations at that point lose weight. express Normalized weights;

[0050] .

[0051] As a further improvement to the combustion state reconstruction method for a combustion engine control system of the present invention:

[0052] The distortion correction loss is:

[0053]

[0054] in, This refers to the number of grid cells in the flame front region and the number of grid cells in the nearby distortion region. Indicates the node position within the reference flame front region. The theoretical physical quantity or the gradient of the theoretical physical quantity at that point. Correcting the loss weights for distortion:

[0055]

[0056] in, This is the adjustment coefficient.

[0057] As a further improvement to the combustion state reconstruction method for a combustion engine control system of the present invention:

[0058] The offline training process of the physical information neural network is as follows:

[0059] (1). For single-condition reconstruction scenarios, all grid nodes corresponding to the target condition in the distorted multiphysics data samples participate in training, and no further division of training / verification / test sets is required;

[0060] For multi-condition generalization scenarios, the samples corresponding to different conditions are randomly divided into training set, validation set and test set in a ratio of 8:1:1. If there is an offline trained physical information neural network for the target condition, the network parameters are directly imported as initial parameters.

[0061] (2). Import offline training data and prior weights of each node, calculate the gradient of the total loss function value with respect to the physical information neural network parameters through backpropagation, update the parameters, and stop training after reaching the preset total number of training steps;

[0062] (3) Evaluate and verify the trained physical information neural network: For single-condition reconstruction scenarios, conduct comparative evaluation with imported data, comparative evaluation with theoretical values, and training convergence evaluation; for multi-condition generalization scenarios, in addition to the three evaluation benchmarks for single-condition reconstruction scenarios, transfer efficiency evaluation and transfer accuracy evaluation are also required.

[0063] As a further improvement to the combustion state reconstruction method for a combustion engine control system of the present invention:

[0064] The comparison and evaluation with the imported data includes: within the high-fidelity region, the mean square error and point-by-point relative error between the calculated physical quantities output by the model and the physical quantities in the distorted multiphysics data samples do not exceed the first preset threshold and the second preset threshold, respectively; within the distortion region, the offset direction of the calculated physical quantities output by the model relative to the center position of the flame front is towards the theoretical physical quantities, and the absolute error between the calculated physical quantities and the theoretical physical quantities is less than the absolute error between the physical quantities in the distorted multiphysics data samples and the theoretical physical quantities;

[0065] The comparison and evaluation with the theoretical value includes: re-obtaining the corrected flame thickness based on the calculated physical quantity, wherein the relative error between the corrected flame thickness and the reference flame thickness is less than a third preset threshold, or the ratio of the absolute error between the corrected flame thickness and the reference flame thickness to the absolute error between the distorted flame thickness and the reference flame thickness is less than a fourth preset threshold.

[0066] The transfer efficiency evaluation includes: the number of training steps required to achieve the same total loss function value after importing similar working condition neural network parameters does not exceed a preset proportion of the number of training steps per working condition when no transfer learning is performed;

[0067] The transfer accuracy assessment includes: after importing neural network parameters for similar operating conditions, the mean square error in the reference flame front region and the high-fidelity region is not greater than a preset multiple of the mean square error when no transfer learning is performed.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] 1. Improve the accuracy of input data in the combustion control system: By reconstructing data in high gradient regions such as the flame front, the temperature field, component concentration field, and flow velocity distribution are obtained that are closer to the real physical process, providing high-confidence input data for mixture ratio control, combustion stability control, and heat load regulation control.

[0070] 2. Real-time reconstruction of combustion state based on sparse observations: This invention can fuse a very small amount of discontinuous experimental measurement data (such as a few PLIF slices or PIV velocity points) to complete the three-dimensional physical field of the entire combustion chamber under strict physical constraints. This provides continuous, high-confidence full-field information for engine real-time condition monitoring systems and fault diagnosis systems, overcoming the limitations of traditional measurement methods.

[0071] 3. Enhanced reliability of the combustion model and generalization ability of the control system: By embedding the combustion control equation as a hard constraint into the neural network, it is ensured that the model output strictly satisfies the physical conservation laws under any operating conditions. This makes the method have good generalization ability under different fuels and different operating conditions. Only the physical property parameters need to be replaced, without retraining the model structure, which facilitates integrated application under varying engine operating conditions and multi-mode operation.

[0072] 4. Support for the full lifecycle development and application of combustion control systems: This invention can be widely applied to multiple stages of combustion control system development and optimization: In the design stage, it is used for high-precision combustion multiphysics calculation and control law simulation verification; in the testing stage, it is used to fuse sparse test data and reconstruct the complete state to evaluate the control effect; in the operation stage, it can be embedded in airborne or ground monitoring systems to provide data support for real-time control and health management.

[0073] 5. Possesses clear engineering application value and rapid portability: The method is based on general control equations, independent of specific CFD software or grids, and can be easily integrated into existing engine design and control processes as a CFD post-processing tool or experimental data fusion module, accelerating the development process of high-performance, high-reliability combustion control systems. Attached Figure Description

[0074] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0075] Figure 1 This is a flowchart illustrating a flame front data reconstruction method based on a physical information neural network according to the present invention.

[0076] Figure 2 This is a schematic diagram of the physical information neural network architecture and loss calculation process of the present invention;

[0077] Figure 3 This is a schematic diagram of a two-dimensional geometric model of the combustion chamber used for CFD simulation calculations;

[0078] Figure 4 A schematic diagram comparing the calculated physical quantities (methane mass fraction) of the flame front using Fluent data and PINN output.

[0079] Figure 5 This is a schematic diagram comparing the calculated physical quantities of the flame front Fluent data, PINN output, and sparse data (methane mass fraction). Detailed Implementation

[0080] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto:

[0081] Example 1: A method for reconstructing combustion state in a combustion engine control system. In the closed-loop control system of a combustion engine, the Electronic Control Unit (ECU) needs to adjust the mixture ratio, control combustion stability, and control heat load distribution based on key state parameters such as the temperature field, component concentration field, reaction zone location, and flow structure within the combustion chamber. Due to the strong abrupt changes in physical quantities in high-gradient regions such as the flame front, traditional numerical simulations or sparse sensor measurements often produce large errors in these regions, thus affecting the control system's accurate judgment of combustion intensity and reaction rate.

[0082] The method described in this embodiment is used to reconstruct and correct the physical consistency of the state in the key high gradient region of the combustion chamber. The obtained temperature field, component mass fraction distribution and multi-physics field information can be used as feedback input parameters of the engine control system for control law calculation and real-time adjustment, thereby improving control accuracy and system stability.

[0083] The specific process of the combustion state reconfiguration method for the combustion engine control system is as follows: Figure 1 As shown, for the engine combustion chamber, distorted computational fluid dynamics (CFD) simulation data is corrected using sparse high-fidelity observation data and physical conservation laws, or a complete multiphysics field conforming to physical laws is obtained using a neural network based on sparse high-fidelity observation data. The steps described in this embodiment are universal and do not depend on specific fuel types or chemical reaction mechanisms; they can be executed simply by replacing the corresponding physical property parameters. Specifically, the steps include:

[0084] Step 1. Data Input and Preprocessing

[0085] In this embodiment, the original multiphysics data of the combustion chamber is obtained through two methods: first, complete multiphysics results obtained by numerical simulation of the target combustion chamber operating conditions using CFD simulation (such as ANSYS Fluent); second, spatially limited measurement data obtained through experimental methods (such as PLIF, PIV, infrared thermography, pressure sensing, etc.). Because CFD simulation is limited by factors such as simplification of turbulence models, simplification of chemical reaction mechanisms, and mesh resolution, its results exhibit significant numerical diffusion and physical distortion in high-gradient regions such as the flame front. Therefore, CFD simulation results are collectively referred to as distorted data. While experimental measurement data has sparse spatial coverage, it has high physical fidelity at the measurement location, hence it is called sparse data, which can be used to provide high-confidence constraints for neural network training. Both are preprocessed as follows:

[0086] (1). Distortion data preprocessing: From the multiphysics results of the combustion chamber obtained by CFD simulation, the discrete multiphysics data at each grid node is extracted according to the computational grid or specified profile. Then, the discrete multiphysics data is normalized and used as a distorted multiphysics data sample to eliminate the influence of dimensions and improve the stability of network training.

[0087] (2) Sparse Data Extension: When the original data source is sparse data obtained from experimental measurements (rather than full-field distorted data from CFD simulation), since the sparse data only covers a limited number of discrete points in the multiphysics field, it cannot be directly used as a sample of distorted multiphysics data covering the entire computational domain. Therefore, it is necessary to first use a temporary pre-trained physical information neural network to extend these sparse points into a physically reasonable initial distribution of the entire field. This temporary pre-trained physical information neural network is only used as a temporary tool and does not require high precision. Its output will be used as distorted data for further correction in subsequent steps. Sparse data includes, but is not limited to: using laser-induced fluorescence (LIF) to measure the concentration distribution of OH* or CH* to determine the location of the reaction zone; using particle image velocimetry (PIV) to obtain velocity field slices; using high-speed infrared thermometry to obtain the temperature of the measurement point; and using pressure sensors / strain gauges to measure the local pressure distribution.

[0088] The specific process for sparse data augmentation is as follows: A physical information neural network is constructed (the specific structure is described in step 3, referring to the physical information neural network architecture). Spatial coordinates are used as input, and physical quantities are used as output. The training of this temporary network uses only two types of constraints: first, sparse data loss, requiring the network output at the measurement point to be consistent with the measured value; second, physical equation loss (such as the residual of the component conservation equation), requiring the network output to satisfy basic physical laws across the entire field. The training of this temporary network uses a conventional optimizer (such as Adam) for a small number of iterations (typically 5000~20000 steps) until the total loss decreases to a plateau (approximately 10). -3 The stopping condition is based on the magnitude of the quantity, and we do not pursue extremely low loss or high precision.

[0089] After inputting sparse data into a temporary pre-trained physical information neural network, its output across all nodes of the computational grid is extracted as the full-field multiphysics distribution. Since this temporary network only accepts sparse constraints and loose physical constraints, its output in high-gradient regions such as flame fronts will inevitably deviate significantly from the actual physical process. Therefore, this full-field output is also considered distorted data, and after normalization, it serves as a distorted multiphysics data sample for further distortion region identification and correction in subsequent steps 2 to 5. This temporary pre-trained network only plays the role of "generating the initial full-field from sparse points" in the entire method; subsequent steps will specifically correct the distorted regions in its output.

[0090] Distorted multiphysics data samples (including preprocessed distorted data, or sparse data expanded using a temporary pre-trained physics information neural network) are used for physical distortion region identification in step 2. Each sample in the distorted multiphysics data sample includes the spatial coordinates of grid nodes. and the corresponding physical quantities physical quantity Including density ,speed ,temperature Mass fraction of each component etc., where i represents the component number, for example, the mass fraction of a component that includes at least methane. Mass fraction of oxygen Mass fraction of carbon dioxide water component mass fraction superscript This indicates normalization.

[0091] Step 2. Regional Division

[0092] To address the problem that traditional PINN technology struggles to automatically identify CFD distortion regions, this method automatically identifies potential distortion regions such as flame fronts by calculating spatial gradients based on key physical quantities, and classifies distorted multiphysics data into high-fidelity data and distorted data.

[0093] Step 2.1. Identification of Physically Distorted Regions

[0094] Physical distortion region identification is based on the spatial gradient of the mass fraction of each fuel component, and the specific process is as follows:

[0095] (1). Identify the center of the flame front

[0096] For the mass fractions of each component obtained in step 1 Calculate the first-order spatial gradient of the mass fraction of fuel components (such as kerosene, methane, hydrogen, etc.) in the fuel mixture. :

[0097] (1)

[0098] in, The grid node position (node ​​coordinates) along the flame normal (or principal gradient direction) is degenerated into axial coordinates in the one-dimensional example to represent the node position; This represents the mass fraction of the fuel components.

[0099] First-order spatial gradient The maximum value is the center of the flame front. For diffusion combustion, the oxidant mass fraction or temperature gradient can also be used as identification variables.

[0100] (2). Normalized gradient field

[0101] (2)

[0102] in, Indicates the mass fraction of fuel components First-order spatial gradient The maximum value, It is used to represent the normalized gradient field.

[0103] (3). Threshold determination

[0104] In multiphysics fields where gradient normalization is applied The spatial location set is determined to be the flame front region, among which... This is a preset gradient threshold, typically ranging from 0.1 to 0.3. The gradient threshold is set... It can automatically identify areas with large gradient changes in the type of flame front. Flame front areas can be regarded as physically distorted areas, and data within these areas can be regarded as physically distorted data.

[0105] Step 2.2. Define the thickness of the distorted flame using the 10%-90% method.

[0106] The flame thickness referred to in this invention specifically refers to the characteristic thickness of a premixed flame (10%–90% method) defined based on the fuel component mass fraction gradient, characterizing the spatial scale of drastic changes in the main physical quantities of the reaction zone. Since the drastic changes in each component actually occur at the same locations, the fuel component mass fraction at the combustion chamber inlet is defined as the fuel component mass fraction in the unburned zone. The mass fraction of fuel components at the combustion chamber outlet is set as the mass fraction of fuel components in the complete combustion zone. ;

[0107] Calculate the 10% position within the range of changes in fuel component mass fraction. The variation range refers to the mass fraction of fuel components in the unburned zone along the flame normal direction. Mass fraction of fuel components in the complete combustion zone The range of variation between these values ​​indicates that the 10% mark represents the mass fraction of fuel components in the unburned zone. Mass fraction of fuel components in the complete combustion zone The spatial location corresponding to a 10% change in the range, the 10% position The mass fraction of the fuel components satisfies:

[0108] (3)

[0109] Similarly, calculate the 90% position within the range of changes in the mass fraction of fuel components. The 90% position refers to the mass fraction of fuel components in the unburned zone. Mass fraction of fuel components in the complete combustion zone The spatial location corresponding to a 90% change in the range between the two, the 90% position The mass fraction of the fuel components satisfies:

[0110] (4)

[0111] The thickness of the distorted flame is: (5)

[0112] The dimensional distortion flame thickness can be obtained by inverse normalization of the distortion flame thickness, or by importing real data and location at the beginning for calculation.

[0113] Step 2.3. Divide the region into high-fidelity region, reference flame front region, and distortion region based on gradient and frontal position.

[0114] The distorted flame thickness is used to further verify whether there is artificial numerical diffusion in the CFD. For distorted data, the value calculated in step 2.2 is used. and The region between the two (i.e. the region corresponding to the thickness of the distorted flame) and the flame front region calculated in step 2.1 are merged, and the region in the combustion chamber space other than the merged region is defined as the high-fidelity region.

[0115] Then, the merged region is further subdivided into a reference flame front region and a general distortion region:

[0116] Reference flame front region: The region centered on the center of the flame front in step 2.1, with a thickness equal to the thickness of the reference flame. The formula for calculating the thickness of the reference flame is:

[0117] (6)

[0118] in, To reference the characteristic gradient of the flame front region, the equilibrium estimation is dominated by the steady-state convection term and the chemical reaction source term:

[0119] (6-a);

[0120] in, To reference the fuel component chemical reaction source term at the center of the flame front region, it was calculated using a one-step overall reaction model—that is, formula (13). , Take the density and normal velocity at the center of the reference flame front region. As a relative reference scale, it is used to delineate the reference flame front region and to determine whether the flame thickness after PINN correction converges in a physically sharper direction, rather than to assert the true physical thickness.

[0121] General distortion area: The part of the merged area excluding the reference flame front area.

[0122] In subsequent steps, unless otherwise specified, "distorted area" refers to the general distortion area (i.e., the part of the merged area that does not include the reference flame front area).

[0123] Each operating condition (i.e., a set of specific initial and boundary conditions, such as a set of values ​​for parameters like inlet velocity, inlet temperature, inlet component mass fraction, and ambient pressure) corresponds to a set of distorted multiphysics data (including any accompanying sparse observation data if available), which is considered a complete sample. Each sample contains the spatial coordinates and normalized physical quantities of all grid nodes under that operating condition, and the identification of physical distortion regions and the division of high-fidelity / reference flame front / distortion regions have been completed according to steps 2.1–2.3.

[0124] Based on practical application requirements, datasets can be constructed in the following two ways:

[0125] (1). Single-condition reconstruction scenario: When the PINN network is only used to correct the distorted multiphysics field under a specific condition, or to complete the full field information based on the sparse observation data of a certain condition, all grid nodes corresponding to the target condition participate in the training, and the training / validation / test set is no longer divided; the offline training of the model is carried out through the process in step 4, and the evaluation and validation are completed through the multi-dimensional comparison with theoretical values ​​and high-fidelity observation data in step 5.

[0126] (2) Multi-condition generalization scenario: When the PINN network needs to learn the multiphysics mapping relationship under multiple conditions simultaneously (in this case, the network input includes control parameters representing the conditions, such as inlet velocity, inlet temperature, and mixing ratio, in addition to spatial coordinates), the samples corresponding to different conditions are randomly divided into training set, validation set, and test set in a ratio of 8:1:1. The training set verifies the generality of each setting of the neural network while training the neural network parameters corresponding to the multiphysics of various conditions. The validation set is used to verify whether the final neural network settings obtained from the training set are appropriate and whether the obtained neural network parameters can meet the transfer learning requirements. The test set is used to finally evaluate the generalization ability of the neural network under unseen conditions. All grid nodes within each sample participate in the calculation of the corresponding stage.

[0127] If the neural network for the target working condition has been built and trained, steps 3 and 4 can be skipped. The samples of the new working condition can be directly imported into the trained neural network for training, and the accuracy of the training results can be evaluated in step 5.

[0128] Similar to step 2.2, the dimensional reference flame thickness and its corresponding position can be obtained by inverse normalization of the flame thickness, or by importing real data and positions at the beginning for calculation. The reference flame front region can be obtained by merging the reference flame thickness and the flame front region obtained in step 2.1. Here, the center position, flame front, and flame tail of the reference flame front can all be obtained together, and the theoretical values ​​corresponding to these positions are directly obtained from the theoretical values ​​of the corresponding positions calculated in step 2.2 for the flame thickness. Only the mass fraction of the fuel components is calculated here; in reality, parameters such as density, velocity, and temperature, as well as the theoretical physical quantities of other components at these positions, can also be calculated using the method in step 2.2. The position of the reference flame front region and the corresponding theoretical physical quantity values ​​at those positions constitute the reference flame front data. Ultimately, we need to make the distorted multiphysics flame front approximate this theoretical value.

[0129] Step 3. Construct PINN

[0130] If this is the first time building a neural network, or if the working conditions for which a neural network is being used for correction differ from those of an existing neural network, then the neural network should be built first. Otherwise, simply import the samples into the neural network for calculation and evaluate the accuracy of the results. The steps for building a neural network are as follows: Build a PINN with spatial coordinates as input and physical quantities as output. The PINN adopts a fully connected deep neural network architecture, including an input layer, 4-8 hidden layers, and an output layer. Each hidden layer contains 20-100 neurons, and the activation function is either tanh or Swish.

[0131] The input to PINN is the spatial coordinates of the node. or (For one-dimensional problems only), time is generally also included. However, this embodiment deals with a steady-state problem, therefore no time variable is introduced. The calculated physical quantities output by PINN include density, temperature, velocity, and the mass fraction of each component.

[0132] Step 4. Construct the loss function and perform offline training of PINN.

[0133] Step 4.1. Construction of the loss function and its weights

[0134] The total loss function of PINN in this invention is a weighted sum of data loss, physical equation loss, and distortion correction loss, specifically in the form of:

[0135] (7)

[0136] in, For data loss, For the weighting coefficients of the data loss term, For the loss of the physical equations, These are the weighting coefficients for the loss term in the physical equation. To correct for distortion loss, These are the weighting coefficients for the distortion correction loss term. The subscript `data` represents the data loss, `phys` represents the physical equation loss, and `low` represents the distortion correction loss. Subsequent parameters also use subscripts with the same meaning.

[0137] (1) Data loss

[0138] (8)

[0139] in, The number of grid points representing data loss includes the high-fidelity region, the reference flame front region, and its vicinity (here, "vicinity" means: taking a range of 3 grid intervals as the distance to the vicinity of the reference flame front region, i.e., assuming the location of the reference flame front region is...). arrive Therefore, the reference area is the flame front region and its vicinity. arrive , among them This refers to the grid spacing (or the average grid spacing within the corresponding range), the total number of grid points in the distorted region, and ns represents the number of computed physical quantities output by the neural network. (Superscript) This indicates normalization, and j represents the node index. The physical information neural network represents the coordinates of the corresponding nodes. The i-th calculated physical quantity output at the point includes density, temperature, velocity, and mass fraction of each component. The node coordinates corresponding to the distorted multiphysics data samples obtained in step 1 The i-th physical quantity at point i, Indicates the coordinates of the node The data loss weight for the i-th physical quantity at a given location. The data loss weight is divided as follows:

[0140] Increase the data loss weights for high-fidelity regions and reference flame front regions to ensure that the calculation results do not deviate from the observation results; decrease the data loss weights for distorted regions to reduce the probability of erroneous regions being learned.

[0141] For the data loss weight of the i-th physical quantity The piecewise function assignment is used, and the formula is as follows:

[0142] (9)

[0143] The data loss weights are extremely large in the high-fidelity region and small in the distorted region. A moderate weight is selected in the reference flame front region. This is to reduce the learning of distorted data while enabling the neural network to fit the high-fidelity data. It is also to prevent the results obtained by the neural network from deviating too much from the imported multiphysics in areas where no modification is needed.

[0144] (2). Physical equation loss

[0145] (10)

[0146] in, This represents the number of grid points lost due to the physical equations, including the total number of grid points in both the distorted and high-fidelity regions. Representing node coordinates The residuals of the component conservation equation at the location, Representing node coordinates The physical equations at that point lose weight because and One-to-one correspondence, therefore and Still a one-to-one correspondence, That is Normalized weights.

[0147] The physical equation loss weights are divided as follows: In distorted regions, the physical quantity distribution is completed using the physical equation loss function, increasing the physical equation loss weight in these regions; in high-fidelity regions, the physical equation loss weight is reduced, as these regions are primarily constrained by data loss. The formula for the physical equation loss weights is as follows:

[0148] (11)

[0149] The role of physical equation loss is to repair the distribution of physical quantities in distorted regions. The physical equation loss weight is larger in distorted regions, while in high-fidelity regions, since data loss has already constrained the data, the physical equation loss weight will be appropriately reduced. At the same time, since the desired result will show a sharp change in gradient near the reference flame front region, the physical equation loss weight of the mesh in the distorted region near the reference flame front region will be relatively small.

[0150] The specific component conservation equations are as follows:

[0151] (12)

[0152] in, For the velocity field, For the first Mass diffusion coefficient of each component For the first The chemical reaction source terms of each component, For the first Evaporation source terms for each component.

[0153] Chemical reaction source term A one-step reaction turnkey model is used to describe this process, and its general form can be expressed as:

[0154] (13)

[0155] in, For the first The stoichiometric coefficients of each component For the first Molar mass of each component The pre-exponential factor of the one-step total reaction is the reaction order parameter; n is the total reaction order of the density term, taken as... sum; As the apparent activation energy, The gas constant is For a local temperature field independent of the composition, Let be the molar masses of the k-th reactant component, respectively. This represents the reaction order (the three parameters here are related to the reactant k, so the subscript is k instead of i). Indicates the number of components participating in the reaction.

[0156] Pre-exponential factors and activation energy The parameters of the one-step turnbuffered reaction model commonly used in the literature, such as the Westbrook–Dryer model or its modified form, can be adopted and calibrated according to the specific fuel type and operating conditions, through experimental data or high-fidelity numerical simulation results.

[0157] For example, consider the methane component in the methane-oxygen reaction:

[0158] (14)

[0159] in, , These are the molar masses of methane and oxygen, respectively. Under this condition, the density in the formula is assumed to be a constant due to its limited range of variation and is integrated into the pre-exponential factor A. In actual implementation, the stoichiometric coefficient is equal to 1 and is therefore omitted. The pre-exponential factor and activation energy here directly use the corresponding values ​​calculated in Fluent. ,activation energy .

[0160] Evaporation source term The volume fraction evaporation model is used to calculate and define the evaporation zone enhancement factor:

[0161] (15)

[0162] in, These are the empirical coefficients for each component. Let be the local liquid volume fraction, where x represents a grid point (location), and t represents the time represented by the local liquid volume fraction at grid point x. The evaporation threshold for component i is defined by an empirical coefficient, and the evaporation threshold itself can be selected based on actual conditions. In this embodiment, the local liquid volume fraction... Obtained from a two-phase spray model. Evaporation threshold. Used to suppress the influence of numerical noise in the thin spray area, its value can be 10. -6 ~10 -4 Select from the range, preferably 10 -5 Empirical coefficient The evaporation enhancement factor is used to adjust the evaporation enhancement amplitude. Its value can be determined by experimental data or high-fidelity numerical results within the range of 0.5 to 5, with an initial value of 1 being preferred. When the combustion system is gas combustion, the evaporation enhancement factor degenerates to a constant of 1.

[0163] when At that time, the evaporation source term is:

[0164] (16)

[0165] in, The basic evaporation rate of the corresponding component can be calculated using the D² law or empirical formulas (such as the Sherwood model).

[0166] Considering that the input to the neural network is normalized coordinates The output consists of normalized parameters. When directly using normalized parameters and coordinates to solve the equation, it is easy to encounter situations where the objective of the residual constraint is not zero. Therefore, when calculating the equation loss, it is necessary to first denormalize the output parameters and then use them uniformly after solving the residuals. To ensure consistency in the solution of the physical equations, the residuals of the equations are normalized to a suitable range. It is assumed that after the normalized physical quantities are denormalized, the magnitudes of the terms in the residuals r(x) are amplified respectively. If the normalized weight is times, then the normalized weight is taken as... This is to constrain the equation residuals to an order of magnitude comparable to other loss terms.

[0167] (3) Distortion correction loss

[0168] (17)

[0169] in, The number of grid points used to correct the distortion loss, including the number of grid points in the flame front region and the number of grid points in the nearby distortion region (similar to data loss, the orientation extends outward by ten grid points as the nearby range). Indicates the node position within the reference flame front region. The theoretical physical quantities at the reference flame front include density, temperature, velocity, and the mass fraction of each component, as well as their theoretical data or gradients. The gradients are derived from the reference flame front data. For data in distorted regions, the imported data corresponding to the leading and trailing edges of the flame front can be used, since theoretically, the vicinity of the reference flame front region corresponds to the theoretical leading and trailing edges of the flame front. Specifically: the theoretical value of the fuel component mass fraction is determined by the 10%-90% method in step 2.2. For other physical quantities such as density, temperature, and velocity, their theoretical values ​​are directly taken from the boundary of the reference flame front region (i.e., the 10% position). and 90% of the positions The corresponding physical quantity values ​​in the distorted multiphysics data samples at the two boundary locations are used as approximations because these two boundary locations are at the junction of the high-fidelity region and the distorted region, where the degree of distortion is relatively low and the data reliability is relatively high. They can be used as reasonable reference values ​​for constraining the distortion correction loss.

[0170] Distortion correction loss weight Adaptively adjust based on gradient magnitude:

[0171] (18)

[0172] in, This is an adjustment coefficient, ranging from 0.5 to 0.9, which makes the weight of regions with larger gradients smaller, thereby preventing extreme losses that are very likely to occur during training from affecting the overall system.

[0173] (4) Prior weights

[0174] For each spatial location in the distorted multiphysics data sample obtained in step 1 Combining the normalized gradient field magnitude at that point calculated in step 2 with the region partitioning result (high-fidelity region, reference flame front region, or distortion region), the spatial location can be obtained using the data loss weight formula, the physics equation loss weight formula, and the distortion correction loss weight formula, respectively. The prior weights are the values ​​of the various loss terms. Because these weights are determined based on the gradient of the imported data and the multiphysics region partitioning, they are called prior weights. These prior weights are used for initial weight assignment during subsequent offline training of the PINN network.

[0175] Step 4.2. Offline training of the PINN network

[0176] The model was trained offline on an NVIDIA GeForce RTX 4080 GPU, and the model parameters were optimized using the Adam optimizer. The initial learning rate of the model was 10. -3 ~10 -2 The decay rate is 0.9 every 5000 steps. The total number of training steps is 10000–50000. Full sampling is used.

[0177] The required offline training data to import includes:

[0178] Based on distorted multiphysics data, for single-condition reconstruction scenarios, import the spatial coordinates of all grid nodes of the sample corresponding to the target condition. and based on the coordinates of each node The corresponding prior weights (i.e., the prior weights of each loss in step 4.1).

[0179] Based on distorted multiphysics data, for multi-condition generalization scenarios, if enough conditions have been trained in the training set or there are samples in the training validation set or test set, then the parameters of the neural network after offline training for similar conditions are imported as initial parameters, so that the neural network obtains a multiphysics field that is basically close to the target during initialization.

[0180] If no PINN trained for the target working condition exists, then import the grid node space coordinates of the samples in the training set. Offline training is performed using corresponding operating parameters (such as inlet velocity and inlet temperature) and prior weights for each node. As the loss decreases during offline training, these prior weights gradually switch to a hybrid gradient calculation method based on both neural network calculation results and imported data, with the weight of the neural network calculation results gradually increasing, up to a maximum of 0.8. Furthermore, regardless of the scenario, physical property parameters such as the pre-exponential factor, activation energy, molar mass of reactant components, and gas constant required for loss calculation in the physical equations must be imported.

[0181] The physical quantities obtained through the output of the PINN network are:

[0182] (19)

[0183] Then, the physical quantities are calculated using Automatic Differentiation (AD). The partial derivatives of each element with respect to spatial coordinates, including the first derivative and the necessary second derivative, are used to calculate the physical equation loss and distortion correction loss. Data loss is calculated directly based on the physical quantities at each point. The total loss is obtained by weighting the various losses. After obtaining the total loss, the gradient of the total loss function value with respect to the neural network parameters is calculated through backpropagation. This gradient is used to update the neural network parameters. Training stops after reaching the preset number of offline training steps, and the parameters of the neural network with the minimum total loss are saved. During offline training, it is easy for the neural network to become overly focused on optimizing certain losses due to their excessive size, thus neglecting other loss terms. Therefore, during training, it is necessary to adjust the basic weights of each loss term and the calculation parameters of the prior weights to achieve better results in offline training of the neural network. Simultaneously, balancing the various losses can be added to ensure that all losses are optimized.

[0184] After completing the offline training in step 4, a basic working neural network can be obtained. However, there are some problems, such as the weight settings for key locations (e.g., the reference flame front and its surroundings) or the calculation method of the loss in actual calculations. These may cause the optimal results obtained by the neural network to not meet the requirements in terms of the accuracy of key locations or calculation results (e.g., the calculated flame thickness). Therefore, it is necessary to continue to optimize the model to obtain a suitable neural network and appropriate parameter settings.

[0185] Step 5. Model Evaluation and Validation

[0186] Step 4 only yields relatively optimized network parameters. The results may still have shortcomings in accuracy or key areas, requiring further evaluation and adjustment of the neural network to obtain an optimized physical information neural network that meets online usage requirements. Evaluation methods vary depending on the application scenario.

[0187] (1) Evaluation of Single-Condition Reconstruction Scenarios: Since the training used all the grid nodes of the scenario, the evaluation does not rely on independent test set partitioning, but is evaluated in multiple dimensions through the following three evaluation benchmarks:

[0188] Comparison and evaluation with imported data: In the high-fidelity region, the mean square error (MSE) and point-by-point relative error between the computed physical quantities output by PINN and the physical quantities in the imported distorted multiphysics data samples are calculated. The mean square error of each physical quantity in the high-fidelity region is required to not exceed a first preset threshold (based on normalized data, preferably 10 in this embodiment). -3Furthermore, the point-by-point relative error does not exceed the second preset threshold (preferably 5% in this embodiment) to ensure that the network does not introduce additional bias in areas where no correction is required; in distorted regions, the difference distribution between the calculated physical quantities output by PINN and the physical quantities in the distorted multiphysics data samples is calculated, and this difference is compared with the theoretical physical quantities of the reference flame front data. For comparison, the calculated physical quantity output by PINN must be offset from the physical quantity in the distorted multiphysics data sample at the center of the flame front in the direction of the theoretical physical quantity value, and the absolute error between the calculated physical quantity and the theoretical physical quantity must be smaller than the absolute error between the physical quantity and the theoretical physical quantity in the distorted multiphysics data sample, that is, the correction magnitude is a positive correction.

[0189] Comparison and evaluation with theoretical values: For key areas (such as the flame front), the method for calculating the CFD flame thickness is exactly the same as in step 2.2. The 10%-90% method is re-applied to the mass fraction distribution output by PINN to obtain the corrected flame thickness. The reference flame thickness obtained in step 2.2 by combining the 10%-90% method with the maximum gradient value For comparison, the relative error of the flame thickness is required to be less than a third preset threshold (preferably 5% in this embodiment) or the absolute error is extremely small. Here, extremely small absolute error means that when the reference flame thickness... In cases where the scale is much smaller than the feature scale of the computational domain, the corrected flame thickness is used. Relative to the thickness of the distorted flame The correction magnitude towards the reference flame thickness value is used as an evaluation metric, requiring the corrected flame thickness to be... and reference flame thickness The absolute error between the two and the distorted flame thickness and reference flame thickness The absolute error ratio between them is less than the fourth preset threshold (preferably 8% in this embodiment), that is .

[0190] For example, in the subsequent experiments in this paper, the reference flame thickness was 0.1 mm, the flame thickness obtained by PINN was 2.8 mm, the absolute flame error was 2.7 mm, the relative error was extremely large, but the accuracy of the absolute error was sufficient for the entire computational domain.

[0191] Training convergence evaluation: Evaluate the convergence of various loss functions (data loss, physical equation loss, distortion correction loss) during the training process, requiring each loss to decrease steadily with training iterations.

[0192] (2) Evaluation of multi-condition generalization scenarios: In addition to the three evaluation benchmarks for single-condition reconstruction scenarios, if the training is not conducted at the beginning of the training set, the following two additional metrics need to be evaluated:

[0193] ① Transfer efficiency evaluation: Record the number of training steps required to reach the same total loss function value after importing neural network parameters for similar working conditions, and compare it with the number of training steps for a single working condition without transfer learning. The former should not exceed 70% of the latter to verify that parameter transfer does indeed compress training time.

[0194] ② Transfer Accuracy Evaluation: The training results of the PINN network trained with parameters from similar operating conditions are compared with the results trained under a single operating condition without transfer learning. Specifically, the mean square error (MSE) between the calculated physical quantities and theoretical physical quantities of PINN within the reference flame front region is calculated under both initialization methods. The MSE after parameter import is required to be no greater than the MSE without transfer learning, meaning the transfer parameters do not amplify the errors in key regions. Simultaneously, the MSE between the calculated physical quantities and the imported distorted multiphysics training samples within the high-fidelity region is calculated under both methods. The MSE after parameter import is required to be no greater than 1.05 times the MSE without transfer learning, ensuring that the transfer parameters do not significantly reduce the overall fitting accuracy.

[0195] If any bias occurs in the above evaluation, the base weights and prior weights need to be adjusted: if there are problems with the data fitting accuracy or convergence, adjust the weights of the data loss and physical equation loss; if the positional error of key locations (such as the flame front) is too large or the flame thickness does not meet the accuracy requirements, check whether the calculation of the reference flame front region is accurate, or investigate whether there are problems with the calculation, balance, or weighting of the distortion correction loss term. In addition, if the problem occurs during transfer learning, it is also necessary to modify the evaluation of similar working conditions or analyze the impact of each initial condition on the multiphysics field to confirm whether the current training conditions are still insufficient.

[0196] After the model is evaluated and validated, the optimal neural network parameters are saved as a physical information neural network optimized to meet the requirements for online use.

[0197] In practical applications, the physical information neural optimization network obtained in step 5, which meets the requirements for online use, is deployed in the engine electronic control unit (ECU) or ground monitoring system. During engine operation, sparse observation data is collected in real time by sensors deployed in and around the combustion chamber. These sensors include, but are not limited to: pressure sensors to obtain local pressure distribution, high-speed infrared thermography to obtain temperature at measurement points, and, under conditions with an optical window, planar laser-induced fluorescence (PLIF) to obtain component concentration slice data or particle image velocimetry (PIV) to obtain velocity field slice data. The collected sparse observation data is normalized according to the sparse data preprocessing method in step 1 and then input into the deployed physical information neural optimization network. The network takes normalized spatial coordinates as input, and after forward inference, outputs normalized physical quantities at each grid node, including density, temperature, velocity, and mass fraction of each component. The normalized physical quantities are then de-normalized to obtain the reconstructed temperature field, component mass fraction distribution, velocity distribution, and flame front position, among other combustion state parameters. The reconstructed combustion state parameters serve as feedback inputs to the control system, used in the calculation of control laws for mixture ratio regulation control, combustion oscillation suppression control, injection flow rate regulation control, and heat load distribution control.

[0198] By improving the accuracy and physical consistency of combustion state perception, this invention can reduce control deviations caused by state estimation errors and improve the stability, response speed and control accuracy of the engine under different operating conditions.

[0199] experiment:

[0200] Experiment 1. Design of Flame Front Based on PINN

[0201] Step 1. Data Input and Preprocessing

[0202] A one-dimensional laminar premixed flame was simulated using ANSYS Fluent 2023 R1. The cross-sectional view of the combustion chamber and the dimensions of each part are shown below. Figure 3 As shown ( Figure 3The dimensions of each part are in millimeters. The methane and oxygen channels consist of two concentric rings. The core of the inner ring is the methane channel, and the outer ring, which forms the annular space between the concentric rings, is the oxygen channel. The radius of the methane channel is 30 mm, and the width of the oxygen channel (i.e., the annular gap width) is 20 mm. The axial distance from the inlet to the starting position of the expansion section at the outlet is 1700 mm, and the total radial dimension at the combustion chamber outlet is 128 mm. The computational domain length is 2200 mm, and the number of mesh nodes at the one-dimensional straight line is 1022. A one-step overall reaction of methane and oxygen is used. The inlet conditions are velocity inlet: methane velocity is 7 m / s, temperature is 313 K, and it enters the combustion chamber from the methane inlet. The mass fraction of methane in this inlet gas stream is 0.9. Oxygen velocity is 34 m / s, temperature is 323 K, and it enters the combustion chamber from the oxygen inlet. The mass fraction of oxygen in this inlet gas stream is 0.767. The export is a pressure export, with an environmental pressure of 101325 Pa.

[0203] Extracting Fluent simulation results, including the node coordinates of 1022 nodes. (j=1,2,...,1022) and the corresponding physical quantity for each node j: methane mass fraction Oxygen mass fraction ,temperature axial velocity and density .

[0204] Then, the methane mass fraction in the physical quantities of each node was calculated. Oxygen mass fraction ,temperature axial velocity ,density Normalization was performed separately:

[0205] (20)

[0206] in, These are the maximum and minimum values ​​of the physical quantity at node j, respectively.

[0207] Step 2. Identification of Physically Distorted Regions

[0208] Calculate the spatial gradient of methane mass fraction at each node j. To avoid inverse normalization, we initially use unnormalized data and locations for the calculation:

[0209] (twenty one)

[0210] Normalize the gradient field:

[0211] (twenty two)

[0212] Set gradient threshold The distorted area is identified as The set of spatial locations, where node j is located within the flame front region, is marked as a distorted data point.

[0213] In this embodiment, the location located A total of 252 nodes within the range are located within the distortion area, which corresponds exactly to the location of the flame front.

[0214] Calculate flame thickness (10%-90% method):

[0215] Unburned methane mass fraction

[0216] methane mass fraction in the complete combustion zone

[0217] 10% position: ,correspond m

[0218] 90% position: ,correspond m

[0219] CFD prediction of flame thickness: m=467mm

[0220] The reference flame thickness is obtained by calculating the center position and gradient of the front: mm.

[0221] It is evident that the flame thickness calculated by Fluent shows significant numerical diffusion compared to the reference flame thickness, which is precisely the distortion area that needs to be corrected.

[0222] Step 3. PINN Network Construction and Training Parameters

[0223] Building a fully connected deep neural network:

[0224] Input layer: 1 neuron, input normalized axial coordinates ;

[0225] Hidden layers: 6 layers, 50 neurons per layer, activation function is ;

[0226] Output layer: 5 neurons, output normalized , , , , .

[0227] Sampling strategy: Full sampling.

[0228] Optimizer: Adam.

[0229] Learning rate decay strategy: First 10,000 iterations: learning rate 0.005; 10,000-20,000 iterations: learning rate decays to 0.0045 (multiplied by 0.9 every 5,000 iterations); 20,000-30,000 iterations: learning rate decays to 0.00405.

[0230] Hardware: Model training was performed on an NVIDIA GeForce RTX 4080 GPU.

[0231] The network was implemented using PyTorch and trained for 30,000 iterations, during which the total loss decreased from the initial 0.2 to [a lower value]. .

[0232] Step 4. Construction of loss function and weights

[0233] The total loss function in this experiment:

[0234] (twenty three)

[0235] in, Used to increase constraints on high-fidelity region data. , Used to prevent excessive losses.

[0236] The data loss, physical equation loss, and distortion correction loss are calculated according to step 4 of Example 1. When calculating the physical equation loss, points far from abrupt gradient changes are selected within the entire computational domain to calculate the component conservation equation residuals. Specifically, since a simple one-dimensional steady-state methane combustion multiphysics field is selected, and atomization is not included, the conservation equations for this example are as follows:

[0237] (twenty four)

[0238] The first and second derivatives are calculated automatically using PyTorch differentiation.

[0239] The chemical reaction source term adopts a one-step turnkey model:

[0240] (25)

[0241] Parameter values: pre-exponential factor ,activation energy molar mass , .

[0242] Step 5. Validation and Evaluation of Training Results

[0243] The trained PINN network is applied to the original 1022 nodes to output the corrected distribution of physical quantities.

[0244] Flame thickness comparison:

[0245] Raw CFD results: mm

[0246] PINN correction results: mm

[0247] Reference flame thickness: mm

[0248] After using the method of this invention, the relative error between the PINN correction result (the calculated physical quantity output by PINN) and the reference flame thickness is extremely large, but the absolute error is already small. The corrected flame thickness... (2.8mm) and reference flame thickness The relative error is relatively large compared to the 0.1mm value, because the reference flame thickness (approximately 0.1mm) is much smaller than the computational domain scale (2200mm). From the correction effect, the distorted flame thickness... The thickness was 467 mm, which was reduced to 2.8 mm after PINN correction. The absolute error between this value and the reference flame thickness was significantly reduced from 466.9 mm to 2.7 mm, a correction of 99.4%, demonstrating a significant positive correction effect and meeting the accuracy requirements. The results of calculating the methane mass fraction and comparing it with the Fluent flame front data are shown in [reference]. Figure 4 In the flame front region, PINN successfully corrected Fluent's numerical diffusion, resulting in a steeper flame structure that more closely resembles the actual physical process. From an engine control application perspective, the flame front position and flame thickness directly affect the control system's judgment of combustion intensity, reaction rate, and heat release rate. Numerical diffusion present in the original CFD results can cause the control system to underestimate local reaction intensity, thereby leading to mixture ratio adjustment deviations or injection flow control errors.

[0249] After correction using the method of this invention, the combustion state parameters are closer to the theoretical values, enabling the control system to more accurately identify the location of the reaction zone and the intensity of combustion, thereby improving the accuracy and response speed of mixture ratio control, combustion stability control and heat load regulation control.

[0250] Experiment 2. Sparse experimental data completion based on PINN

[0251] This example is used to verify the effectiveness of the present invention in engine control state reconstruction. For sparse data that is likely to be encountered in actual situations, the sparse data can be imported into the neural network and trained, and the result can be imported as Fluent data for re-front correction. In this experiment, the neural network result from Experiment 1 is selected as the baseline, and 40 data points are selected as sparse data for multiphysics completion and compared with the Fluent result.

[0252] Step 1. Refer to step 3 of Experiment 1 to build a neural network, and use the data fitting term and physical loss term as loss to train the neural network to obtain distorted multiphysics data;

[0253] Step 2. Import the multiphysics data obtained from the neural network in Step 1 as distorted data, and repeat the entire process of Experiment 1 to obtain the completed multiphysics data. See [link to methane mass fraction completion] for details. Figure 5 .from Figure 5 As can be seen, the red dashed line (control group, CFD simulation, i.e., Fluent calculation result) represents a relatively gradual decline and does not fit the high-fidelity observation data represented by the blue circle. However, the PINN network of this invention successfully completes the multiphysics field based on a small amount of data, making the completed multiphysics field closer to the actual physical process than the Fluent result. This demonstrates the effectiveness of the sparse multiphysics data completion technique used in this invention and shows significant technological advancements in the calculation of key high-gradient regions compared to CFD simulation. In actual engine operation, sensors typically provide only limited local measurement data, making it difficult to directly obtain the complete combustion state field. The method of this invention can reconstruct the complete combustion state field based on a small amount of high-confidence observation data, enabling the control system to obtain continuous and physically consistent state feedback information even under conditions of limited sensor deployment.

[0254] This capability significantly enhances the robustness and stability of the engine control system under complex operating conditions and with incomplete sensor data, and improves the real-time adjustment capability of the closed-loop control system for the combustion process.

[0255] Finally, it should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for reconstructing combustion state in a combustion engine control system, characterized in that... Includes the following steps: S1. Obtain distorted multiphysics data samples of the combustion chamber. The distorted multiphysics data samples include the spatial coordinates and physical quantities of the grid nodes. The physical quantities include at least the mass fraction of fuel components. S2. Determine the center position of the flame front based on the spatial gradient of the fuel component mass fraction, determine the thickness of the distorted flame based on the variation of the fuel component mass fraction, and then divide the spatial region of the combustion chamber into the reference flame front region, the distorted region and the high-fidelity region. S3. Construct a physical information neural network and a total loss function that includes data loss, physical equation loss, and distortion correction loss; assign prior weights to the loss of each grid node according to whether it belongs to the high-fidelity region, the reference flame front region, or the distortion region. S4. Use distorted multiphysics data samples and prior weights to train the physical information neural network offline to obtain the optimized physical information neural network; S5. Real-time acquisition of sparse observation data from the combustion chamber, input into the physical information neural optimization network, and output of reconstructed physical quantities.

2. The combustion state reconstructing method for a combustion engine control system according to claim 1, characterized in that: The method for acquiring distorted multiphysics data samples in step S1 includes: (1). If the original data obtained is obtained by numerical simulation of the target combustion chamber operating conditions through CFD simulation, then the discrete multiphysics data at each grid node is extracted and then normalized. (2). If the original data is obtained from sensor measurements, sparse data expansion is first performed using a temporary pre-trained physical information neural network that includes data loss and physical equation loss, and then normalization is performed.

3. The combustion state reconstructing method for a combustion engine control system according to claim 2, characterized in that: The process of dividing the reference flame front region, the distortion region, and the high-fidelity region in step S2 includes: (1) Calculate the first-order spatial gradient of fuel component mass fraction at each grid node. , Indicates the position of the mesh node along the flame normal; (2). First-order spatial gradient The location of the maximum value is determined as the center of the flame front; (3). The first-order spatial gradient Normalization is performed to obtain the normalized gradient field. The set of spatial locations in the normalized gradient field that are greater than the preset gradient threshold is taken as the flame front region. (4). The region centered on the center of the flame front and with a thickness equal to the reference flame thickness is taken as the reference flame front region; the region corresponding to the distorted flame thickness is merged with the flame front region, and the high-fidelity region is the region outside the merged region. The part of the merged region other than the reference flame front region is taken as the distorted region.

4. The combustion state reconstructing method for a combustion engine control system according to claim 3, characterized in that: The thickness of the distorted flame satisfy: ; in, This represents the 10% mark within the range of variations in the mass fraction of fuel components. This represents the 90% position within the range of changes in the mass fraction of fuel components. The reference flame thickness satisfies: ; in, Indicates the reference flame thickness. For reference, the characteristic gradient of the flame front region: ; in, To reference the fuel component chemical reaction source term at the center of the flame front region, , For reference, the density and velocity at the center of the flame front region; and Fuel component mass fractions at the 10% and 90% positions, respectively: , ; in, The mass fraction of fuel components in the unburned zone. This represents the mass fraction of fuel components at the combustion chamber outlet.

5. The combustion state reconstructing method for a combustion engine control system according to claim 4, characterized in that: The data loss is: ; in, This represents the total number of grid points in the high-fidelity region, the reference flame front region, and the distortion region within three adjacent grid areas. ns represents the number of computed physical quantities output. Indicates the coordinates of the node The i-th calculated physical quantity output at point i. Represents the node coordinates in a distorted multiphysics data sample. The i-th physical quantity at point i, Indicates the coordinates of the node The data loss weight for the i-th physical quantity at point i: ; in, This represents the normalized gradient field.

6. The combustion state reconstructing method for a combustion engine control system according to claim 5, characterized in that: The loss of the physical equation is: ; in, This represents the total number of grid points in the distorted region and the high-fidelity region. Representing node coordinates The residuals of the component conservation equation at the location, Representing node coordinates The physical equations at that point lose weight. express Normalized weights; 。 7. The combustion state reconstructing method for a combustion engine control system according to claim 6, characterized in that: The distortion correction loss is: ; in, This refers to the number of grid cells in the flame front region and the number of grid cells in the nearby distortion region. Indicates the node position within the reference flame front region. The theoretical physical quantity or the gradient of the theoretical physical quantity at that point. Correcting the loss weights for distortion: ; in, This is the adjustment coefficient.

8. The combustion state reconstructing method for a combustion engine control system according to claim 7, characterized in that: The offline training process of a physical information neural network is as follows: (1). For single-condition reconstruction scenarios, all grid nodes corresponding to the target condition in the distorted multiphysics data samples participate in training, and no further division of training / verification / test sets is required; For multi-condition generalization scenarios, the samples corresponding to different conditions are randomly divided into training set, validation set and test set in a ratio of 8:1:

1. If there is an offline trained physical information neural network for the target condition, the network parameters are directly imported as initial parameters. (2). Import offline training data and prior weights of each node, calculate the gradient of the total loss function value with respect to the physical information neural network parameters through backpropagation, update the parameters, and stop training after reaching the preset total number of training steps; (3) Evaluate and verify the trained physical information neural network: For single-condition reconstruction scenarios, conduct comparative evaluation with imported data, comparative evaluation with theoretical values, and training convergence evaluation; for multi-condition generalization scenarios, in addition to the three evaluation benchmarks for single-condition reconstruction scenarios, transfer efficiency evaluation and transfer accuracy evaluation are also required.

9. A method for reconstructing the combustion state in a combustion engine control system according to claim 8, characterized in that: The comparison and evaluation with the imported data includes: within the high-fidelity region, the mean square error and point-by-point relative error between the calculated physical quantities output by the model and the physical quantities in the distorted multiphysics data samples do not exceed the first preset threshold and the second preset threshold, respectively; within the distortion region, the offset direction of the calculated physical quantities output by the model relative to the center position of the flame front is towards the theoretical physical quantities, and the absolute error between the calculated physical quantities and the theoretical physical quantities is less than the absolute error between the physical quantities in the distorted multiphysics data samples and the theoretical physical quantities; The comparison and evaluation with the theoretical value includes: re-obtaining the corrected flame thickness based on the calculated physical quantity, wherein the relative error between the corrected flame thickness and the reference flame thickness is less than a third preset threshold, or the ratio of the absolute error between the corrected flame thickness and the reference flame thickness to the absolute error between the distorted flame thickness and the reference flame thickness is less than a fourth preset threshold. The transfer efficiency evaluation includes: the number of training steps required to achieve the same total loss function value after importing similar working condition neural network parameters does not exceed a preset proportion of the number of training steps per working condition when no transfer learning is performed; The transfer accuracy assessment includes: after importing neural network parameters for similar operating conditions, the mean square error in the reference flame front region and the high-fidelity region is not greater than a preset multiple of the mean square error when no transfer learning is performed.

Citation Information

Patent Citations

  • CNN-PINN-based gas turbine combustion chamber performance prediction method

    CN121211618A

  • Multi-region reaction network modeling method for combustion chamber of heavy-duty gas turbine

    CN122117076A