A three-dimensional flow field reconstruction method for a hydraulic torque converter for sparse data

Abstract

This invention relates to the field of fluid machinery testing and intelligent flow field evaluation technology, and provides a method for reconstructing the three-dimensional flow field of a hydraulic torque converter for sparse data. This method uses a small number of known cross-sections and sparse internal observation points as anchor points to construct a physical information neural network. In model training, in addition to conventional data fitting and fluid dynamics equation residuals, a variable cross-section volumetric flow rate consistency constraint based on the three-dimensional geometry of the flow channel is introduced, and preheating annealing and an adaptive weight balancing mechanism are used to achieve a synergistic reduction of multiple loss functions. This method effectively improves the mass conservation bias of conventional reconstruction methods under the condition of missing complete inlet and outlet boundaries, achieving physical consistency of the three-dimensional flow field, and is suitable for the analysis and evaluation of the flow field within a hydraulic torque converter.

Description

Technical Field

[0001] This invention belongs to the fields of fluid mechanics and artificial intelligence, and relates to a method for reconstructing the three-dimensional flow field of a hydraulic torque converter for sparse data. Background Technology

[0002] The internal flow channels of complex turbomachinery such as hydraulic torque converters are typically variable cross-section, exhibiting complex turbulent characteristics such as nonlinearity, large local gradients, and three-dimensional vortices. While traditional computational fluid dynamics methods can provide high-resolution flow field details, they rely on complete inlet and outlet boundary conditions and require high-quality mesh generation. In engineering practice, this makes it difficult to fully meet the need for rapid reconstruction of the internal characteristics of the flow field.

[0003] In practical flow field analysis of complex turbomachinery, the inlet and outlet boundary conditions of the flow channels, such as guide wheels, are often difficult to extract completely due to the complex three-dimensional structure with variable cross-sections and the difficulty of data extraction. Furthermore, the extracted observation data is often sparse and discontinuous; typically, only two-dimensional flow field observation data of a partial cross-section and a small number of internal discrete points can be obtained in engineering practice. This lack of complete inlet and outlet boundaries and the scarcity of effective data make directly reconstructing a continuous three-dimensional flow field an ill-conditioned inverse problem.

[0004] Existing methods generally suffer from the following problems when dealing with complex variable cross-section flow channels such as hydraulic torque converters, where inlet and outlet boundary conditions are lacking:

[0005] (1) Lack of physical consistency: Traditional spatial interpolation algorithms are based on geometric relationships. On the one hand, they lack fluid dynamics constraints, and on the other hand, they are difficult to guarantee the continuity of cross sections and the consistency of flow, resulting in physical distortion of the reconstruction results.

[0006] (2) Physical violation problem: Pure data-driven deep neural networks are prone to overfitting under sparse data, and the output results often violate the basic laws of fluid physics, exhibiting eddies that violate physical laws and lack interpretability.

[0007] (3) Insufficient macroscopic constraints: Existing Physics-Informed Neural Networks (PINN) frameworks mostly rely on supervised learning of complete boundaries and are limited by local physical partial differential equation residual constraints. They lack effective integration of three-dimensional geometric features of the flow channel, such as introducing global physical constraints like cross-sectional volume flow conservation.

[0008] Therefore, there is an urgent need for a method that can stably reconstruct the three-dimensional flow field of complex variable cross-section channels and improve the physical consistency of the reconstruction results under conditions where complete inlet and outlet boundaries are missing and effective data extraction is limited. Summary of the Invention

[0009] The purpose of this invention is to provide a method for reconstructing the three-dimensional flow field of a hydraulic torque converter for sparse data, aiming to solve the problems mentioned in the background art.

[0010] The present invention is implemented as follows: a method for reconstructing the three-dimensional flow field of a hydraulic torque converter for sparse data includes the following steps:

[0011] Step S1: Determine the target three-dimensional spatial subdomain. The spatial boundary of the target three-dimensional spatial subdomain is formed by the known upstream cross-section, the known downstream cross-section, and the solid wall surface formed by the blade surface, the inner annular surface of the flow channel, and the outer annular surface of the flow channel. Obtain the three-dimensional coordinate data of the discrete points of the global fluid space within the target three-dimensional spatial subdomain. The training of the model and the reconstruction of the flow field are both limited to this three-dimensional spatial subdomain.

[0012] Step S2: Obtain flow field data of multiple known cross-sections within the target three-dimensional spatial subdomain, and obtain flow field data of sparse internal observation points extracted from discrete points in the global fluid space; wherein, the multiple known cross-sections include at least the upstream known cross-section constituting the spatial boundary, the downstream known cross-section, and intermediate cross-sections located within the subdomain.

[0013] Step S3: Construct the PINN model. The PINN model takes spatial coordinates as input and outputs at least three-dimensional velocity components, pressure, and turbulent kinematic viscosity.

[0014] Step S4: Construct a joint loss function, which includes at least: a data loss term for fitting the flow field data of known cross-sections and sparse internal observation points, a boundary loss term for applying solid wall constraints, an equation residual loss term for introducing physical constraints of the fluid dynamics control equations, and a variable cross-section volumetric flow rate consistency loss term.

[0015] Step S5: Arrange multiple constraint cross-sections along the main flow direction within the target three-dimensional spatial subdomain, and obtain the cross-sectional area of ​​each constraint cross-section to construct a variable cross-section volumetric flow rate consistency constraint; wherein, the constraint cross-section is used to apply macroscopic flow rate conservation conditions, and its spatial location includes at least one of the following: a position independent of the known cross-section and a position partially overlapping with the known cross-section.

[0016] Step S6: The PINN model is trained using a multi-stage collaborative training strategy. This strategy includes dynamically adjusting the weight coefficients of each term in the joint loss function. The dynamic adjustment is based at least in part on the statistics of the backpropagation gradient of each loss term, so as to achieve the transition of first fitting the flow field data and then gradually introducing physical and flow constraints. The network weights are updated through the backpropagation of the joint loss function to obtain the target subdomain flow field reconstruction result that satisfies the cross-sectional flow conservation and physical consistency.

[0017] In a further technical solution, in step S2, the flow field data of multiple known cross-sections includes at least three-dimensional velocity component data.

[0018] In a further technical solution, in step S2, sparse internal observation points are distributed within the target three-dimensional spatial subdomain, and their total number is composed of global basic sampling points and local gradient densification points; wherein, the number of basic sampling points is positively correlated with the volume of the target three-dimensional spatial subdomain, and the number of local gradient densification points is positively correlated with the spatial local gradient amplitude of the flow field variables.

[0019] A further technical solution involves generating the spatial distribution of sparse internal observation points within the target's three-dimensional spatial subdomain using Latin hypercube sampling.

[0020] In a further technical solution, in step S5, the cross-sectional area of ​​each constraint surface is obtained by at least one of the following methods: pre-extracting the geometric features of the solid wall of the flow channel entity and calling it as a geometric constant term during PINN model training; or dynamically integrating based on the current geometric boundary of the cross-section in a single iteration of PINN model training.

[0021] In a further technical solution, in step S6, the weight coefficients of the equation residual loss term and the volumetric flow rate consistency loss term are adjusted in stages according to the training process: in the initial stage of training, the weight coefficients are maintained at a preset initial value; after the preset training progress is reached, the weight coefficients increase with the increase of the number of iteration steps until the preset target range is reached.

[0022] A further technical solution is to set the upper limit threshold for the weight of the volumetric flow rate consistency loss term higher than the upper limit threshold for the weight of the equation residual loss term.

[0023] The present invention provides a method for reconstructing the three-dimensional flow field of a hydraulic torque converter for sparse data, the beneficial effects of which are as follows:

[0024] (1) Reduce boundary dependence: Under the condition of missing complete inlet and outlet boundaries, a small number of known cross sections and sparse internal observation points are used to effectively reconstruct the three-dimensional flow field in the subdomain, thereby reducing the dependence on complete boundary conditions in flow field analysis.

[0025] (2) Improve physical consistency: By introducing a variable cross-section volumetric flow consistency loss term, the local fluid dynamics equation residuals are combined with the macroscopic cross-section flow conservation constraints, which improves the mass conservation deviation that is prone to occur in the variable cross-section flow channel by conventional data-driven algorithms or pure PINN methods, thereby improving the physical interpretability and credibility of the reconstruction results.

[0026] (3) Improve data utilization: By using a multi-stage training strategy of preheating annealing and adaptive weight balancing, the divergence problem that occurs during the collaborative optimization of multiple constraint terms is alleviated, and the convergence stability and reconstruction accuracy of the model under sparse data conditions are further improved.

[0027] (4) Convenient implementation: Prior information such as the geometric area of ​​the variable cross section of the flow channel can be pre-calculated offline, the process is highly standardized, and it is convenient for engineering applications. Attached Figure Description

[0028] Figure 1 A flowchart of a three-dimensional flow field reconstruction method for a hydraulic torque converter oriented towards sparse data, provided in an embodiment of the present invention;

[0029] Figure 2 A schematic diagram illustrating the data input, network architecture, and multiple joint loss functions of the PINN model;

[0030] Figure 3 A schematic diagram of the three-dimensional solid geometric model and flow field mesh of the variable cross-section flow channel of the hydraulic torque converter guide wheel (where a is the three-dimensional solid geometric model and b is the flow field mesh).

[0031] Figure 4 This is a schematic diagram of the three-dimensional spatial distribution of three known cross-sections and discrete observation points within the target subdomain.

[0032] Figure 5 A schematic diagram of the variable cross section taken for the consistency constraint of volumetric flow rate and a curve showing the change of cross-sectional area (where a is a schematic diagram of the variable cross section and b is a curve showing the change of cross-sectional area).

[0033] Figure 6 The graph shows the convergence curves of various loss functions during model training (where a is the data loss term, b is the boundary loss term, c is the equation residual loss term, and d is the volumetric flow rate consistency loss term).

[0034] Figure 7 This is a comparison diagram of the flow field reconstruction distribution in the target subdomain using the method of this embodiment, the spatial interpolation method, and the reference results (where a is the method of this invention, b is the spatial interpolation method, and c is the CFD simulation).

[0035] Figure 8 This is a comparison diagram of the residual distribution in the reconstructed flow field between the method of the present invention and the spatial interpolation method;

[0036] Figure 9 This is a comparison chart of the volumetric flow rate variation trends between cross sections along the mainstream direction using the method of this invention and the spatial interpolation method. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0038] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0039] like Figure 1 As shown, an embodiment of the present invention provides a method for reconstructing the three-dimensional flow field of a hydraulic torque converter for sparse data, comprising the following steps:

[0040] Step S1: Determine the target three-dimensional spatial subdomain. The spatial boundary of the target three-dimensional spatial subdomain is formed by the known upstream cross-section, the known downstream cross-section, and the solid wall surface formed by the blade surface, the inner annular surface of the flow channel, and the outer annular surface of the flow channel. Obtain the three-dimensional coordinate data of the discrete points of the global fluid space within the target three-dimensional spatial subdomain. The training of the model and the reconstruction of the flow field are both limited to this three-dimensional spatial subdomain.

[0041] Step S2: Obtain flow field data of multiple known cross-sections within the target three-dimensional spatial subdomain, and obtain flow field data of sparse internal observation points extracted from discrete points in the global fluid space; wherein, the multiple known cross-sections include at least the upstream known cross-section constituting the spatial boundary, the downstream known cross-section, and intermediate cross-sections located within the subdomain.

[0042] Step S3: Construct the PINN model. The PINN model takes spatial coordinates as input and outputs at least three-dimensional velocity components, pressure, and turbulent kinematic viscosity.

[0043] Step S4: Construct a joint loss function, which includes at least: a data loss term for fitting the flow field data of known cross-sections and sparse internal observation points, a boundary loss term for applying solid wall constraints, an equation residual loss term for introducing physical constraints of the fluid dynamics control equations, and a variable cross-section volumetric flow rate consistency loss term.

[0044] Step S5: Arrange multiple constraint cross-sections along the main flow direction within the target three-dimensional spatial subdomain, and obtain the cross-sectional area of ​​each constraint cross-section to construct a variable cross-section volumetric flow rate consistency constraint; wherein, the constraint cross-section is used to apply macroscopic flow rate conservation conditions, and its spatial location includes at least one of the following: a position independent of the known cross-section and a position partially overlapping with the known cross-section.

[0045] Step S6: The PINN model is trained using a multi-stage collaborative training strategy. This strategy includes dynamically adjusting the weight coefficients of each term in the joint loss function. The dynamic adjustment is based at least in part on the statistics of the backpropagation gradient of each loss term, so as to achieve the transition of first fitting the flow field data and then gradually introducing physical and flow constraints. The network weights are updated through the backpropagation of the joint loss function to obtain the target subdomain flow field reconstruction result that satisfies the cross-sectional flow conservation and physical consistency.

[0046] In this embodiment of the invention, the method can be specifically applied to the reconstruction and performance evaluation of at least one flow field selected from the guide wheel, pump wheel and turbine of a hydraulic torque converter.

[0047] In a preferred embodiment of the present invention, in step S2, the flow field data of multiple known cross-sections includes at least three-dimensional velocity component data.

[0048] In a preferred embodiment of the present invention, in step S2, sparse internal observation points are distributed within the target three-dimensional spatial subdomain, and their total number is composed of global basic sampling points and local gradient densification points; wherein, the number of basic sampling points is positively correlated with the volume of the target three-dimensional spatial subdomain, and the number of local gradient densification points is positively correlated with the spatial local gradient magnitude of the flow field variables.

[0049] In a preferred embodiment of the present invention, the spatial distribution of sparse internal observation points within the target three-dimensional spatial subdomain is generated using Latin hypercube sampling to avoid local spatial aggregation and achieve uniform coverage of the three-dimensional fluid domain.

[0050] In a preferred embodiment of the present invention, in step S5, the cross-sectional area of ​​each constraint surface is obtained by at least one of the following methods: pre-extracting the geometric features of the solid wall of the flow channel entity and calling it as a geometric constant term during the training of the PINN model; or dynamically integrating based on the geometric boundary of the current surface in a single iteration of the PINN model training.

[0051] In a preferred embodiment of the present invention, in step S6, the weight coefficients of the equation residual loss term and the volumetric flow rate consistency loss term are adjusted in stages as the training process progresses: in the initial stage of training, the weight coefficients are maintained at a preset initial value; after the preset training progress is reached, the weight coefficients increase with the increase of the number of iteration steps until the preset target range is reached.

[0052] In a preferred embodiment of the present invention, the upper limit threshold of the weight of the volumetric flow consistency loss term is greater than the upper limit threshold of the weight of the equation residual loss term. This is used to enhance the cross-sectional macroscopic flow conservation constraint while suppressing the excessive suppression of the data loss term by the equation residual loss term in the local high gradient region, so that the model optimization process is guided by the data loss term.

[0053] The following are some specific examples to verify the effectiveness of this method.

[0054] Example 1: Flow field reconstruction of the guide wheel subdomain based on "3 known cross-sections + 500 internal discrete points";

[0055] For the guide wheel subdomain of the hydraulic torque converter, 500 internal observation points were selected. The number 500 was not arbitrarily set, but rather the number of feature points calculated based on sampling principles. The specific process is as follows:

[0056] The basic sampling base is established: First, based on the three-dimensional volume of the guide wheel subdomain and combined with the spatial Nyquist sampling criterion, a preliminary judgment is made. In order to capture the overall flow field information in the flow channel, the number of basic sampling points to maintain the residual constraints of the global physical equation is calculated to be 300. These 300 points ensure that no vacuum zone of physical information appears in the smooth mainstream region.

[0057] Feature gradient refinement: By performing preliminary low-precision flow field prediction on this guide vane subdomain, high gradient regions such as blade trailing edge shedding vortices and suction surface boundary layers are identified. Based on the local velocity gradient magnitude, approximately 200 additional refinement sampling points are allocated in this complex flow region to capture local flow field characteristics.

[0058] Balancing computational power and accuracy: Based on the combined results of the two calculations above, the total number of internal discrete observation points was ultimately determined to be 500. Testing showed that with this point configuration, the model achieved an optimal balance between reconstruction accuracy and computational convergence speed.

[0059] The next step is to reconstruct the flow field:

[0060] 1. Definition of a subdomain: In dimensionless coordinates, combined with... Figure 3 a and Figure 3 b) Three known cross-sections are selected to form a target three-dimensional spatial subdomain, and discrete points of the global fluid space within this subdomain are extracted as the input coordinates of the network. The training and parameter evaluation of the PINN model are both limited to this subdomain, without introducing the flow field outside the subdomain or complete inlet and outlet boundary data.

[0061] 2. Composition of Observational Data: The flow field data consists of two parts (see...). Figure 4 (1) Flow field data from three known cross sections, mainly consisting of velocity components; (2) 500 sparse internal observation points extracted from the entire space within the target subdomain to supplement internal spatial information. Inlet and outlet boundary data are not included in the supervision, simulating a reconstruction scenario where inlet and outlet boundary conditions are missing.

[0062] 3. Model and joint loss function construction: such as Figure 2 As shown in the figure, , , Represents the coordinates of three known cross-sections. , , This represents the coordinates of 500 known internal points. , , Represents the coordinates inside and at the boundary of the fluid domain. , , (This refers to the coordinates input to the network after dimensionless processing.) This method uses the PINN model, with three-dimensional spatial coordinates as input. The output includes at least a three-dimensional velocity component vector U=(u, v, w) (where u, v, and w are the velocity components of the fluid along the x, y, and z axes, respectively), and pressure. And turbulence-related quantities. Specifically, turbulence-related quantities include turbulent kinematic viscosity. (A non-negative activation function is used to ensure its physical meaning).

[0063] Constructing a joint loss function :

[0064]

[0065] in, For data loss items, For boundary loss terms, This is the residual loss term in the equation. This is the volumetric flow rate consistency loss term. , , as well as The dynamic weight coefficients representing the four losses mentioned above are updated in real time through warm-up annealing and adaptive mechanisms during network training.

[0066] The specific definitions of the above four types of losses are as follows:

[0067] (1) Data loss items :

[0068] The formula used to fit the flow field data of three known cross sections and sparse interior observation points, constraining the consistency between the network predictions and known observations, is as follows:

[0069]

[0070] In the formula, The three-dimensional velocity predicted by the network model. This is the actual observed velocity vector extracted through physical measurement methods.

[0071] (2) Boundary loss term :

[0072]

[0073] In the formula, This is the velocity vector predicted by the model at the solid wall boundary of the flow channel, thereby constraining the fluid velocity at the wall to approach 0.

[0074] (3) Equation residual loss term :

[0075] Physical constraints are introduced to the Reynolds-averaged Navier-Stokes (RANS) equations. The residuals of the RANS momentum equation and continuity equation are embedded into a network to achieve a macroscopic physical fit to complex turbulent flow fields. The calculation formula is as follows:

[0076]

[0077] In the formula, ν is the Reynolds number (dimensionless). t Let be the kinematic viscosity (dimensionless), and ∇ be the vector differential operator. The sum of squares in the first term of the formula corresponds to the residual of the momentum conservation equation under the Boussinesq eddy viscosity assumption, which includes convection terms. Pressure gradient term and viscous stress terms The second term corresponds to the mass conservation (continuity equation) residual, i.e. .

[0078] (4) Volumetric flow rate consistency loss item :

[0079] This is the core macroscopic constraint of this method, used to ensure the conservation of volumetric flow rate across the cross-section within the variable cross-section channel. Its calculation formula and the method of introducing the geometric area will be explained in detail in the following steps.

[0080] 4. Variable cross-section volumetric flow rate consistency constraint: Within the target subdomain, 10 constraint cross-sections are arranged along the main fluid flow direction, such as... Figure 5 a (where x represents the coordinate of the main flow direction of the channel, and y and z represent the coordinates of the cross-sectional plane perpendicular to the main flow direction) and Figure 5 b (where, This is the difference between the largest and smallest cross-sectional areas among the 10 constrained cross-sections. The average area of ​​the 10 constraint cross-sections is shown in the figure. The geometric area of ​​each constraint cross-section is obtained by preprocessing the geometric features of the solid flow channel model and is introduced as a known parameter during the training phase of the PINN model.

[0081] No. Constrained cross-sectional volume flow rate The definition of is:

[0082]

[0083] in, For the first Cross-sectional area This represents the average normal velocity of the tangential surface.

[0084] Volumetric flow rate consistency loss term Defined as:

[0085]

[0086] In the formula, The total number of constraint surfaces arranged within the target subdomain. For the first Constrained cross-sectional volume flow rate.

[0087] 5. Training Strategy: The training process employs a multi-stage collaborative strategy of preheating and soft annealing to dynamically adjust the weight coefficients of the aforementioned joint loss function. , , as well as ):

[0088] The warm-up phase occurs in the early stages of model training. and Set to zero or a minimum value to prioritize model optimization. and This stage prompts the network-predicted flow field to be anchored to known observation points and the solid boundary of the flow channel in order to construct the initial flow field framework.

[0089] The soft annealing stage and the adaptive weighting module are gradually integrated after the warm-up period. and To mitigate convergence oscillations caused by differences in gradient magnitudes in joint optimization due to multiple physical and data constraints, an adaptive weight module (Self-Adaptive Loss, SAL) is introduced to dynamically update the weights of each loss term.

[0090] During backpropagation of the network, the gradient magnitude gi of each loss term with respect to the shared hidden layer parameters is dynamically calculated. To avoid drastic fluctuations in gradients during a single iteration, an exponential moving average (EMA) is used to smooth historical gradient information. Its calculation formula is as follows:

[0091]

[0092] In the formula, This represents the current training iteration step. This is the number of steps in the previous training iteration. The translational slip attenuation coefficient, This is the gradient magnitude statistic for each loss term in the current training iteration step. This refers to the smoothed gradient features from the previous training iteration. This represents the smoothed gradient feature quantity. Based on... The network dynamically calculates and updates. , , as well as This ensures that the gradient contributions provided by each loss term during backpropagation tend to be balanced. Simultaneously, an upper bound truncation mechanism is introduced to... A mandatory numerical upper limit threshold is set to prevent the residual loss term from excessively suppressing the sparse data supervision term due to a surge in numerical values ​​in local high gradient regions. Furthermore, it allows... Maintain a high upper limit to preserve the conservation constraint of cross-sectional macroscopic flow throughout the entire training cycle.

[0093] 6. Reconstruction Results: After training according to the above process, the convergence curves of each loss term are as follows: Figure 6 a- Figure 6 As shown in Figure d, continuous three-dimensional velocity and pressure field reconstruction results are obtained within the target subdomain. The results indicate that:

[0094] In the absence of import and export supervision, combined with Figure 7 a and Figure 7 As shown in Figure c, the flow field reconstruction distribution comparison diagram shows that the main flow structure in the subdomain can be stably reconstructed.

[0095] Compared to schemes that rely solely on interpolation, the physical consistency metrics are superior, such as... Figure 8 As shown.

[0096] like Figure 9 The contrasting line graphs showing the volumetric flow rate change trends indicate that the flow rate fluctuations across the cross section have decreased, and the consistency of the volumetric flow rate has improved.

[0097] Therefore, it can be seen that under the limited conditions of "a small number of cross-sections + sparse internal points + no inlet and outlet boundaries", this method can achieve the reconstruction of complex flow channel subdomains with physical consistency.

[0098] Comparative Example 1: Subdomain reconstruction method based solely on interpolation;

[0099] Under the same data conditions as in Example 1 (same known cross section and same 500 internal discrete points), without using the PINN model, without introducing consistency constraints on equation residuals and volumetric flow rate, only the interpolation method is used to reconstruct the subdomain flow field.

[0100] 1. Comparative method: Spatial interpolation of the velocity at the observation points yields the predicted velocity field on the subdomain grid. This method does not include: residual constraints of the governing equations, consistency constraints of variable cross-section volumetric flow rate, and multi-loss collaborative training mechanisms.

[0101] 2. Comparison Results: Example 1 and Comparative Example 1 were compared on the same evaluation cross section and the same evaluation mesh. The results show:

[0102] Depend on Figure 7 b and Figure 7 As shown in c, Comparative Example 1 achieves a smoother local visualization; however, combined with... Figure 9 It can be seen that Comparative Example 1 shows large fluctuations in cross-sectional flow consistency indicators (such as the dispersion of volumetric flow rate Q);

[0103] like Figure 8 The residual distribution comparison chart shown indicates that the continuity residual index (residuals of the continuity equation and momentum equation) of Comparative Example 1 is significantly worse than that of Example 1.

[0104] Example 1 demonstrates superior overall performance in balancing reconstruction error and physical consistency.

[0105] 3. Comparative conclusions: Comparative example 1 shows that although interpolation alone can obtain locally smooth results, it is difficult to guarantee the conservation consistency and physical interpretability in complex variable cross-section flow channels. This method, by synergistically integrating data constraints, equation constraints and volumetric flow consistency constraints, can achieve more reliable subdomain flow field reconstruction under the condition of missing inlet and outlet boundaries.

[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for reconstructing three-dimensional flow field of a hydraulic torque converter for sparse data, characterized in that, Includes the following steps: Step S1: Determine the target three-dimensional spatial subdomain. The spatial boundary of the target three-dimensional spatial subdomain is formed by the known upstream cross-section, the known downstream cross-section, and a solid wall surface composed of the blade surface, the inner annular surface of the flow channel, and the outer annular surface of the flow channel. Obtain the three-dimensional coordinate data of discrete points of the global fluid space within the target three-dimensional spatial subdomain. The training of the model and the flow field reconstruction are both limited to this three-dimensional spatial subdomain. Step S2: Obtain flow field data of multiple known cross-sections within the target three-dimensional spatial subdomain, and obtain flow field data of sparse internal observation points extracted from discrete points in the global fluid space; wherein, the multiple known cross-sections include at least the upstream known cross-section, the downstream known cross-section, and the intermediate cross-section located within the subdomain that constitute the spatial boundary. Step S3: Construct the PINN model. The PINN model takes spatial coordinates as input and outputs at least three-dimensional velocity components, pressure, and turbulent kinematic viscosity. Step S4: Construct a joint loss function, which includes at least: a data loss term for fitting the flow field data of known cross-section and sparse internal observation points, a boundary loss term for applying solid wall constraints, an equation residual loss term for introducing physical constraints of the fluid dynamics control equations, and a variable cross-section volumetric flow rate consistency loss term. Step S5: Arrange multiple constraint cross-sections along the main flow direction within the target three-dimensional spatial subdomain, and obtain the cross-sectional area of ​​each constraint cross-section to construct a variable cross-section volumetric flow rate consistency constraint; wherein, the constraint cross-section is used to apply macroscopic flow rate conservation conditions, and its spatial location includes at least one of the following: a position independent of the known cross-section and a position partially overlapping with the known cross-section; Step S6: The PINN model is trained using a multi-stage collaborative training strategy. This training strategy includes dynamically adjusting the weight coefficients of each term in the joint loss function. The dynamic adjustment is at least partially based on the statistics of the backpropagation gradient of each loss term. The flow field data is fitted first, and then physical and flow constraints are introduced. The network weights are updated through the backpropagation of the joint loss function to obtain the target subdomain flow field reconstruction result that satisfies the cross-sectional flow conservation and physical consistency.

2. The method of claim 1, wherein, In step S2, the flow field data of multiple known cross-sections includes at least three-dimensional velocity component data.

3. The method of claim 2, wherein, In step S2, sparse internal observation points are distributed within the target three-dimensional spatial subdomain, and their total number is composed of global basic sampling points and local gradient densification points. Among them, the number of basic sampling points is positively correlated with the volume of the target three-dimensional spatial subdomain, and the number of local gradient densification points is positively correlated with the spatial local gradient magnitude of the flow field variables.

4. The method of claim 3, wherein, The spatial distribution of sparse internal observation points within the target's three-dimensional spatial subdomain is generated using Latin hypercube sampling.

5. The method for reconstructing the three-dimensional flow field of a hydraulic torque converter for sparse data according to claim 1, characterized in that, In step S5, the cross-sectional area of ​​each constraint surface is obtained in at least one of the following ways: pre-extracted based on the geometric features of the solid wall of the flow channel and called as a geometric constant term during PINN model training; or obtained by dynamic integration based on the geometric boundary of the current surface in a single iteration of PINN model training.

6. The method for reconstructing the three-dimensional flow field of a hydraulic torque converter for sparse data according to claim 1, characterized in that, In step S6, the weight coefficients of the equation residual loss term and the volumetric flow rate consistency loss term are adjusted in stages as the training process progresses: in the initial stage of training, the weight coefficients are maintained at a preset initial value. After the preset training progress is reached, the weight coefficients increase with the number of iterations until the preset target range is reached.

7. The method for reconstructing the three-dimensional flow field of a hydraulic torque converter for sparse data according to claim 6, characterized in that, The upper limit threshold for the weight of the volumetric flow rate consistency loss term is greater than the upper limit threshold for the weight of the equation residual loss term.

Images