A high-precision prediction method for a hydrodynamic torque converter cavitation flow field

Abstract

The present application belongs to the technical field of intersection of computational fluid dynamics and artificial intelligence, and in particular to a high-precision prediction method for a hydrodynamic torque converter cavitation flow field. The method comprises the following steps: S1: obtaining transient CFD flow field simulation data inside the guide wheel flow passage of the hydrodynamic torque converter; S2: constructing a multi-task neural network model; S3: constructing a hybrid supervised loss function for optimizing the cavitation prediction sub-network; S4: training the neural network model by adopting a three-stage decoupling strategy; S5: constructing a physical residual loss term based on the law of conservation of fluid dynamics; and S6: inputting the space-time coordinates to be predicted into the trained neural network model. The present application improves the recognition accuracy of sparse cavitation features, enhances the training stability of complex multiphase flow fields, improves the physical consistency of the prediction results, and significantly improves the calculation efficiency of the cavitation flow field.

Description

Technical Field

[0001] This invention relates to the field of computational fluid dynamics and artificial intelligence, specifically to a high-precision prediction method for cavitation flow field of a hydraulic torque converter. Background Technology

[0002] The hydraulic torque converter is a core component of automotive automatic transmissions (AT). Under extreme conditions such as stall, its internal flow exhibits extremely complex unsteady characteristics. In particular, vortex cavitation flow induces cavitation in the near-wall region of the blades due to a sudden drop in static pressure, forming a complex flow field structure with localized gas-liquid coexistence. This phenomenon is particularly pronounced in the near-wall region of the guide vane blades. The occurrence of cavitation not only causes instability in the near-wall flow field structure of the blades, altering the internal flow state, but also directly affects the energy conversion efficiency of the hydraulic torque converter. Simultaneously, the cavitation effect generated during cavitation collapse can cause material erosion on the blade surface, thereby inducing fatigue failure risks. Currently, the analysis of the cavitation flow field in hydraulic torque converters mainly relies on CFD numerical simulation technology based on the Navier-Stokes equations. However, due to the strong nonlinearity and multiphase coupling characteristics of cavitation flow, high-precision transient CFD simulations often face problems such as high computational resource requirements and long solution cycles, limiting their engineering applications in airfoil optimization and real-time operating condition assessment.

[0003] In recent years, with the rapid development of artificial intelligence technology, deep learning-based methods for rapid flow field prediction have gradually attracted attention. In pure data-driven modeling, researchers have widely adopted architectures such as Convolutional Neural Networks (CNNs) and U-Nets (U-shaped Networks) to construct surrogate models, achieving efficient prediction of airfoil flow and steady-state flow fields, demonstrating good inference performance.

[0004] While the aforementioned methods perform well under specific conditions, their direct application to predict cavitation flow fields in complex fluid machinery such as hydraulic torque converter guide vanes still faces significant challenges. First, under extreme conditions, the cavitation flow field evolution in the guide vane region is drastic, with rapid transient flow responses and sparse effective samples. Conventional models tend to overlook local phase transition characteristics, leading to predictions that degenerate into pure liquid flow and lose physical meaning. Second, the dramatic step changes in volume fraction at the gas-liquid interface make gradient propagation instable in regression models, limiting the accuracy of numerical fitting. Furthermore, purely data-driven methods lack physical constraints, making it difficult to guarantee that prediction results satisfy fundamental physical laws such as mass conservation, and they also exhibit poor adaptability to large curvature boundaries of guide vane blades. Therefore, there is an urgent need to develop a rapid cavitation flow field prediction method that balances sample imbalance handling, boundary recognition capabilities, and physical consistency constraints, thereby promoting the development of forward design theory and performance evaluation methods for hydraulic torque converter blades. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0007] A high-precision prediction method for cavitation flow field in a hydraulic torque converter includes the following steps:

[0008] S1: Obtain transient CFD flow field simulation data inside the guide wheel channel of the hydraulic torque converter, extract the spatiotemporal coordinates in the three-dimensional fluid domain of the hydraulic torque converter guide wheel, as well as the corresponding discrete flow field physical quantities and the original gas phase volume fraction; construct a binary cavitation label characterizing the cavitation region distribution based on a preset physical threshold.

[0009] S2: Construct a multi-task neural network model that includes a shared backbone network, a flow field prediction subnetwork, and a cavitation prediction subnetwork;

[0010] S3: Construct a hybrid supervised loss function for optimizing the cavitation prediction subnetwork;

[0011] S4: A three-stage decoupling strategy is used to train the neural network model:

[0012] Freeze the cavitation prediction subnetwork and pre-train the shared backbone and flow field prediction subnetwork using flow field physical quantity data;

[0013] The frozen flow field prediction subnetwork utilizes binary cavitation labels and liquid phase volume fraction, and centrally optimizes the hybrid supervised loss function based on a cavitation sensing sampling strategy.

[0014] Unfreeze all network parameters and perform multi-task joint fine-tuning;

[0015] S5: In the joint fine-tuning stage, a physical residual loss term based on the conservation law of fluid dynamics is constructed and added to the total loss function; the partial derivatives of the predicted flow field with respect to the spatiotemporal coordinates are calculated using automatic differentiation technology, and the model output is constrained to meet the conditions of mass conservation and momentum conservation.

[0016] S6: Input the spatiotemporal coordinates to be predicted into the trained neural network model to obtain the cavitation confidence score; map the confidence score to the cavitation probability through the Sigmoid activation function, and reconstruct the continuous liquid phase volume fraction distribution of the entire flow field as the final prediction result of the cavitation flow field.

[0017] As a preferred embodiment of the high-precision prediction method for the cavitation flow field of a hydraulic torque converter described in this invention, the CFD simulation in S1 is based on the Mixture multiphase flow model, wherein the cavitation phase change source term is calculated using the Schnerr-Sauer model; the evaporation rate in the Schnerr-Sauer model... With condensation rate The calculation formula is as follows:

[0018]

[0019]

[0020] in, This represents the mass source term generated by liquid vaporization. This represents the mass source term generated by gas condensation; The density of the liquid phase is... For gas phase density, The density of the mixed phase; The preset nucleation volume fraction, Where is the bubble radius; Let be the saturated vapor pressure of the fluid at the current temperature. This represents the local pressure in the flow field.

[0021] As a preferred embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention, wherein a binary cavitation label is constructed in step S1. The specific method is as follows: Let the actual liquid volume fraction obtained from CFD simulation be... The physical threshold for determining cavitation is set as follows: , The value range is from 0.8 to 0.95; for any sample point Its binary emptying label Defined as:

[0022]

[0023] in, This indicates that the point is located in a cavitation region. This indicates that the point is located in a non-cavitation region.

[0024] As a preferred embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention, in S2, the shared backbone network is used to extract high-dimensional implicit features of normalized spatiotemporal coordinates; the flow field prediction subnetwork is used to output velocity field and pressure field based on the implicit features; and the cavitation prediction subnetwork is used to output unnormalized cavitation confidence based on the implicit features.

[0025] The neural network model employs the Softplus activation function to ensure the continuity of higher-order derivatives when calculating physical residuals in S5. The Softplus activation function is defined as follows:

[0026]

[0027] In step S6, the scalar cavitation confidence score output by the cavitation prediction subnetwork is... Mapped to vacuolation probability and the final predicted liquid volume fraction The formula is:

[0028]

[0029]

[0030] in, This represents the Sigmoid function. The range of values ​​is .

[0031] As a preferred embodiment of the high-precision prediction method for cavitation flow field of hydraulic torque converter described in this invention, the hybrid supervision loss function in S3 consists of three weighted components: a classification loss term based on the binary cavitation label, a regression correction loss term calculated only in the real cavitation region, and a false positive penalty term calculated only in the real non-cavitation region.

[0032] The hybrid supervision loss function From classification loss terms Regression Correction Loss Term and false positive penalty items The weighted composition, its mathematical expression is:

[0033]

[0034] in, These are the classification loss weight coefficients; These are the regression loss weight coefficients; This represents the weighting coefficient for false positive penalties.

[0035] As a preferred embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention, the classification loss term... The focus loss function is used, and the calculation formula is as follows:

[0036]

[0037] in, To focus parameters, the weights of easily distinguishable samples are reduced so that the model focuses on difficult-to-distinguish samples; This is a balancing parameter used to adjust for the imbalance between positive and negative samples; The probability of the model predicting the true class is defined as:

[0038] .

[0039] As a preferred embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention, the regression correction loss term Only applicable to specific cavitation point sets with real labels The calculation uses the mean absolute error form; false positive penalty term. Only for point sets whose real labels are non-empty Calculation; the specific formula is as follows:

[0040]

[0041]

[0042] in, This represents the set of samples from the real cavitation region. This represents the set of samples from the real non-empty regions; It is a linear rectification function, only when the predicted value Below the threshold Punishment will be imposed at that time.

[0043] As a preferred embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention, the physical residual loss term in S5 is... Including the residuals of the continuity equation for incompressible fluids With momentum equation residual Its definition is as follows:

[0044]

[0045]

[0046] The formula for calculating the total physical residual loss is:

[0047]

[0048] in, The velocity vector output by the flow field prediction subnetwork. To predict static pressure, To incorporate the mixed-phase fluid density calculated using cavitation phase change, The dynamic viscosity of the mixed-phase fluid. For the nabla operator, This represents the total number of sample points used to calculate the physical residual.

[0049] As a preferred embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention, the cavitation sensing sampling strategy in S4 specifically involves dividing the training dataset into cavitation sample sets. Non-empty sample set When constructing each training batch, the sampling ratio of cavitation samples is forcibly set to [value missing]. , If the batch size is Then from Randomly selected One sample, from Randomly selected There are samples that satisfy:

[0050]

[0051]

[0052] in, This indicates rounding a numerical value to the nearest integer; proportion This is used to force the model to acquire sufficient cavitation feature gradients in each iteration step.

[0053] As a preferred embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter described in this invention, the parameter update logic of the three-stage evolutionary training in S4 is as follows: Let the shared backbone network parameter set be... The parameter set of the flow field prediction subnetwork is The parameter set of the cavitation prediction subnetwork is Phase 1 involves freezing the cavitation prediction subnetwork, pre-training the shared backbone and flow field prediction subnetwork using flow field physical quantity data, and then updating only a subset of parameters in the optimizer. To minimize Phase two involves loading the parameters trained in phase one. The frozen flow field prediction subnetwork utilizes binary cavitation labels and liquid phase volume fraction, and centrally optimizes the hybrid supervised loss function based on a cavitation-sensing sampling strategy. The optimizer only updates a subset of parameters. To minimize Phase three involves unfreezing all network parameters, performing multi-task joint fine-tuning, and the optimizer jointly updating all parameters to minimize the total loss. .

[0054] Compared with existing technologies, the beneficial effects of this invention are: 1. Improved recognition accuracy of sparse cavitation features: Addressing the sample imbalance problem caused by the small proportion of cavitation regions in hydraulic torque converters, this invention employs a hybrid supervision mechanism of "classification-led, regression-corrected" combined with a cavitation sensing sampling strategy. This method effectively captures cavitation boundaries through binary classification, uses regression correction to fit the liquid phase volume fraction gradient within the cavitation region, and uses a false positive penalty term to suppress false alarms in non-cavitation regions, thus improving the phenomenon that conventional methods easily overlook sparse cavitation features.

[0055] 2. Enhanced training stability for complex multiphase flow fields: The three-stage evolutionary training strategy proposed in this invention optimizes model parameters step by step in the order of "flow field pre-training - cavitation-specific training - joint fine-tuning". This strategy effectively decouples the learning process of flow field features and cavitation features, alleviates optimization conflicts caused by differences in gradients of different physical quantities in multi-task learning, and helps the model achieve stable convergence in complex phase change flow field tasks.

[0056] 3. Improved physical consistency of prediction results: This invention introduces physical residual constraints based on fluid dynamics governing equations (such as the Navier-Stokes equations and continuity equations) during the joint model fine-tuning stage. By embedding physical conservation laws into the loss function through automatic differentiation technology, the purely data-driven prediction results are corrected, making the predicted flow field more consistent with physical laws and improving the model's generalization ability in the case of partial data loss.

[0057] 4. Significantly improved computational efficiency of cavitation flow fields: The neural network model trained by this invention enables rapid flow field inference. Compared with traditional CFD numerical simulation of flow fields, this method significantly shortens the computation time while ensuring a certain level of prediction accuracy, making it suitable for rapid parameter scanning and performance evaluation in the design phase of hydraulic torque converters. Attached Figure Description

[0058] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0059] Figure 1 This is a flowchart of TSML-ChainPINN cavitation training and prediction for a high-precision prediction method of cavitation flow field of a hydraulic torque converter according to the present invention.

[0060] Figure 2 This is a mesh model diagram of the flow channel of the hydraulic torque converter guide wheel in an embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention;

[0061] Figure 3 This is a three-stage evolutionary training flowchart in an embodiment of a high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention.

[0062] Figure 4 This is a diagram of the TSML-ChainPINN architecture in an embodiment of a high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention.

[0063] Figure 5 This is a diagram showing the extraction section position of the guide vane of the hydraulic torque converter in an embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention.

[0064] Figure 6 This is a comparison and error diagnosis diagram of cavitation binary masks in an embodiment of a high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention.

[0065] Figure 7 This is a comparison diagram of the liquid phase volume fraction focused in the cavitation region in an embodiment of the high-precision prediction method for the cavitation flow field of a hydraulic torque converter according to the present invention.

[0066] Figure 8 This is an example of a high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention, showing CFD, PINN baseline model, TSML-ChainPINN three-dimensional cavitation prediction and error distribution diagram;

[0067] Figure 9 This is a TSML-ChainPINN three-stage loss function curve in an embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention.

[0068] Figure 10 This is a scatter plot showing the correlation between the predicted flow velocity and the CFD baseline value in an embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention.

[0069] Figure 11 This is a scatter plot showing the correlation between the predicted pressure value and the CFD reference value in an embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention.

[0070] Figure 12 This is a box plot comparing the model performance of multiple physical quantity global evaluation indexes in an embodiment of the high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention.

[0071] Figure 13 This is a comparison of the computational efficiency of traditional CFD and TSML methods in an embodiment of a high-precision prediction method for cavitation flow field of a hydraulic torque converter according to the present invention. Detailed Implementation

[0072] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0073] Secondly, the present invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include three-dimensional spatial dimensions of length, width, and depth.

[0074] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0075] Example

[0076] This embodiment uses the prediction of cavitation flow field of guide vane under stall conditions of a certain type of hydraulic torque converter as an example to illustrate the specific implementation process of the method described in this invention. The computing hardware platform used in this embodiment is a workstation equipped with a high-performance graphics processing unit (GPU), and the software environment includes Ansys Fluent 2024R2, ParaView 6.01, and a deep learning program developed based on the PyTorch 2.0 framework.

[0077] Please see Figure 1 A high-precision prediction method for cavitation flow field in a hydraulic torque converter includes the following steps:

[0078] Step S1: Data Acquisition and Hollowing Tag Construction

[0079] The core of this step lies in constructing a high-fidelity numerical computing environment to obtain the basic data required for neural network training and to complete label construction. First, this invention uses industry-recognized high-fidelity transient CFD numerical simulation results as the training data source and model performance evaluation benchmark. These simulation results have undergone rigorous mesh independence verification and numerical method effectiveness verification. Using 3D modeling software such as SpaceClaim or UG, a single-channel fluid domain model of the hydraulic torque converter guide vane is extracted through Boolean operations. Subsequently, the geometric model is imported into ICEMCFD or AnsysMeshing for mesh generation. High-quality hexahedral or polyhedral meshing techniques are employed, with a focus on boundary layer refinement of the guide vane blade suction surface and leading edge region. The initial mesh height is set to 0.01 mm, the layer growth rate is 1.2, and the total number of boundary layers is 15, ensuring that the dimensionless wall distance y+ value is strictly controlled between 1 and 5. The final generated mesh has approximately 310,000 nodes. Figure 2As shown. After mesh generation, the mesh model was imported into the Ansys Fluent solver for numerical calculation settings. A Mixture multiphase flow model coupled with a Schnerr-Sauer cavitation model was used, and the SSTk-omega model, suitable for strong shear flows, was selected for the turbulence model. Boundary conditions were set with an inlet velocity of 18 m / s to simulate stall conditions, an outlet as a pressure outlet, and unsteady calculation mode was enabled with a time step of 0.0005 s. After the flow field evolved to a quasi-periodic fluctuation state, intensive sampling was started and transient calculation results were saved. After data acquisition, the transient result file generated by Fluent was read using the visualization software ParaView. Key physical field variables, including three-dimensional velocity vectors, static pressure, and gas phase volume fraction, were extracted using the software's built-in calculation filters. The flow field data for each time step was then exported as a Visualization Toolkit (VTK) format file using batch processing. Finally, the VTK files of all time steps are read in the Python programming environment, and the coordinates of the sample points and their corresponding physical quantities are extracted. Based on the mean and standard deviation of the training set, the Z-Score Normalization method is used to normalize the input coordinates and target physical quantities. Based on the physical conservation relationship of multiphase flow, the gas phase volume fraction is converted into the liquid phase volume fraction. At the same time, a physical threshold of 0.95 is introduced to construct a binary classification label. That is, for each spatial sample point, if its liquid phase volume fraction is less than 0.95, it is marked as a cavitation region; otherwise, it is marked as a non-cavitation region, thus completing the conversion from continuous field to classification label.

[0080] Step S2: Construct a chained multi-task neural network model

[0081] This step involves building a TSML-ChainPINN neural network model based on the PyTorch deep learning framework, which includes a shared backbone and multi-task branches. Figure 3As shown, a shared backbone network is first constructed as the feature extractor. This network is designed as a multilayer perceptron with 10 fully connected layers, and each layer has 512 neurons. To meet the requirements of higher-order derivatives in subsequent physical residual calculations, the inter-layer activation function is the Softplus function (β=2.0), which has infinitely differentiable characteristics. Following the shared backbone network, two task-specific branches are connected in parallel to achieve feature decoupling. The flow field prediction sub-network consists of four fully connected layers, responsible for outputting normalized velocity and pressure vectors. Its output layer has no activation function to allow continuous regression in the real number domain. The cavitation prediction sub-network also consists of four fully connected layers, with its terminal specifically designed to directly output a scalar Logits value instead of a probability value. This Logits value will subsequently be mapped to a cavitation probability using a Sigmoid activation function and converted into the final predicted liquid volume fraction through a complementary relationship. This chain-like structure design allows the model to simultaneously handle continuous flow field regression and discrete cavitation classification tasks while sharing spatiotemporal flow field features.

[0082] Step S3: Define the hybrid supervised loss function

[0083] This step defines a three-part hybrid supervised total loss function to address the challenges of cavitation sample sparsity and gradient instability. The first part is the classification loss term, which uses FocalLoss to calculate the error between the predicted Logits and the binary labels. The focusing parameter γ is set to 2.0, and the balancing parameter α is set to 0.25. This increases the model's attention to difficult-to-classify samples such as cavitation initiation and collapse boundaries while reducing the weight of a large number of simple non-cavitation samples. The second part is the regression correction loss term, which employs a conditional calculation strategy and is activated only within the sample set whose true label is cavitation. It uses L1 norm constraints to predict the liquid volume fraction to approximate the true value, enabling the network to further fit the numerical gradient of the cavitation core region after identifying the cavitation region. The third part is the false positive penalty term, which is calculated only in regions where the true label is non-cavitation. It specifically applies a squared penalty to sample points that are incorrectly predicted as cavitation (i.e., predicted values ​​below a threshold), effectively suppressing false noise in non-cavitation regions. The three losses are weighted and summed according to preset weight coefficients (e.g., classification loss term: regression correction loss term: false positive penalty term = 1.0: 1.0: 2.0) to form the final optimization objective.

[0084] Step S4: Perform three-stage evolutionary training

[0085] This step uses the Adam optimizer to perform rigorous, phased, decoupled training on the model, such as... Figure 4As shown. The first stage is flow field pre-training, where the parameters of the cavitation prediction head are frozen and cannot be updated. Only the shared backbone and the flow field prediction head are trained using flow field physical quantity data, with approximately 80 epochs of iteration. The aim is to enable the shared backbone network to prioritize mastering the basic topology and spatiotemporal evolution of the flow field. The second stage is cavitation-specific training, where the flow field prediction head parameters are frozen, the cavitation prediction head is unfrozen, and a cavitation-aware sampling mechanism is introduced. When constructing each batch, an index is used to force the inclusion of 30% cavitation samples and 70% non-cavitation samples. The hybrid supervised loss function is optimized centrally and iterated for approximately 120 epochs to address the prediction problem caused by sample imbalance. The third stage is joint fine-tuning, where all network parameters are unfrozen, allowing the weights of the entire network to be fine-tuned under the combined effect of multi-task gradients, preparing for the introduction of physical constraints.

[0086] Step S5: Introduce physical residual constraints

[0087] This step introduces physical constraints based on the conservation laws of fluid dynamics during the joint fine-tuning phase to correct non-physical predictions that may arise from purely data-driven approaches. Specifically, using the automatic differentiation engine built into the PyTorch framework, the velocity and pressure outputs of the flow field prediction subnetwork are differentiated with respect to the input spatiotemporal coordinates to construct the residuals of the incompressible Navier-Stokes equations and the continuity equations. The continuity residuals calculate the velocity divergence, constraining the flow field to satisfy mass conservation; the momentum residuals encompass unsteady terms, convection terms, pressure gradient terms, and viscous diffusion terms, constraining the flow field to satisfy momentum conservation. The mean square sum of the above residuals is defined as the physical loss term, and this physical loss term is weighted with a weight of 0.1 and added to the multi-task total loss function. This ensures that the neural network not only fits the observed data but also conforms to the fundamental governing equations of fluid dynamics at the derivative level, thereby significantly improving the model's generalization ability and physical reliability in sparse data regions.

[0088] Step S6: Cavitation Flow Field Reasoning and Reconstruction

[0089] This step describes the inference and field reconstruction process of the model after training. The spatiotemporal coordinates of the test set are fed into the trained neural network model as input. The model outputs flow field variables and cavitation logits simultaneously through forward propagation. For cavitation prediction, the model's output logits are first mapped to cavitation probabilities between 0 and 1 using a sigmoid activation function, and then transformed into a continuous liquid volume fraction distribution through a linear transformation. Finally, using the mean and standard deviation statistically derived from the training set, all normalized prediction outputs are denormalized to restore velocity, pressure, and liquid volume fraction with real physical units, thereby reconstructing a fine transient cavitation morphology and flow field structure across the entire flow field.

[0090] Step S7: Evaluation and Visualization of Prediction Results

[0091] This step involves multi-dimensional quantitative evaluation and visualization analysis of the reconstructed flow field. The predicted data is exported to VTK format, and ParaView or Tecplot software is used for flow field visualization post-processing. A representative guide wheel mid-section (12mm from the inner ring surface) is selected as the two-dimensional analysis section to conduct two-dimensional cavitation flow field characteristic analysis. The results are as follows: Figure 5 As shown. The cavitation boundary with a liquid volume fraction of 0.95 was extracted using isosurface analysis, as shown... Figure 6 , Figure 7 As shown, the dark area represents the cavitation core region, and the light area represents the pure liquid phase region. The liquid phase volume fraction error cloud map shows that the prediction error is concentrated in the thin layer region at the cavitation boundary, with the magnitude controlled within 10. -3 Level; combined with two-dimensional voiding mask ( Figure 6 ) and three-dimensional cavitation morphology ( Figure 8 Qualitative error analysis was conducted. Traditional PINN baseline models suffer from false positives (FP), false negatives (FN), and numerous discrete spurious noise at cavitation boundaries. This invention effectively suppresses non-physical divergences caused by extreme sample imbalance, reduces false positives and false negatives, eliminates spurious noise, and provides cavitation boundary predictions that more closely match the CFD ground truth. It also achieves high-fidelity 3D reconstruction of sheet-like cavitation and detached cloud morphology details, exhibiting a higher degree of overlap with the ground truth. In contrast, baseline models suffer from over-smoothing and local distortions. Figure 9 As shown, the loss function curve indicates that the total loss and sub-losses such as data loss and physical loss continuously decrease and converge during the three-stage training of the model, without significant oscillations or divergence, demonstrating good stability and convergence of the training framework; for the pressure field and velocity field, scatter plots ( Figure 10 , Figure 11 Quantitative analysis showed that, compared to the PINN baseline model, the TSML-ChainPINN model exhibited improved coefficient of determination (R²) for the velocity field (from 0.8624 to 0.9813), decreased root mean square error (RMSE) from 4.27 m / s to 1.58 m / s, and reduced mean absolute error (MAE) from 1.89 m / s to 0.650 m / s; for the pressure field, R² increased from 0.9823 to 0.9985, and RMSE decreased from 2.72 × 10⁻⁶. 4 Pa decreased to 7.85 × 10³ Pa, and MAE decreased from 1.89 × 10 4 With Pa reduced to 3.71 × 10³ Pa, the predicted points better match the ideal y = x line, and the dispersion is significantly reduced. For example... Figure 12As shown, the box plot represents the sample distribution characteristics of VOF F1 score, VOF Intersection over Union (IoU), mean absolute error of pressure (MAE), and root mean square error of velocity (RMSE). Compared with the PINN baseline model, the TSML-ChainPINN model reduces pressure MAE and velocity RMSE by approximately 79.8% and 62.8%, respectively. The boxes and whiskers of both indices shrink significantly, resulting in stronger consistency and robustness among samples. The VOF F1 score increases from 0.621 to 0.918, an improvement of approximately 47.8%; the VOF IoU increases from 0.455 to 0.848, an improvement of approximately 86.4%. The baseline model exhibits greater overall dispersion in VOF F1 score, VOF IoU, pressure MAE, and velocity RMSE, with a few outliers appearing in pressure MAE and velocity RMSE. The VOF index boxes of this invention are more compact, demonstrating significant advantages in prediction stability and achieving a net benefit of balancing multiple physical quantities. Figure 13 As shown, the inference time per time step of this invention is approximately 0.53 seconds, achieving a computational speedup of about 157 times compared to traditional CFD unsteady solutions, while maintaining an extremely low cumulative time consumption, demonstrating a significant advantage in computational efficiency. Experimental results show that this invention achieves a synergistic improvement in computational accuracy and solution efficiency while ensuring prediction accuracy and physical consistency, providing a reliable surrogate model scheme for the efficient design and optimization iteration of hydraulic torque converters.

[0092] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A high-precision prediction method for cavitation flow field in a hydraulic torque converter, characterized in that, Includes the following steps: S1: Obtain transient CFD flow field simulation data inside the guide wheel channel of the hydraulic torque converter, extract the spatiotemporal coordinates in the three-dimensional fluid domain of the guide wheel of the hydraulic torque converter, as well as the corresponding discrete flow field physical quantities and the original gas phase volume fraction; A binary cavitation label representing the distribution of cavitation regions is constructed based on a preset physical threshold. S2: Construct a multi-task neural network model that includes a shared backbone network, a flow field prediction subnetwork, and a cavitation prediction subnetwork; S3: Construct a hybrid supervised loss function for optimizing the cavitation prediction subnetwork; S4: A three-stage decoupling strategy is used to train the neural network model: Freeze the cavitation prediction subnetwork and pre-train the shared backbone and flow field prediction subnetwork using flow field physical quantity data; The frozen flow field prediction subnetwork utilizes binary cavitation labels and liquid phase volume fraction, and centrally optimizes the hybrid supervised loss function based on a cavitation sensing sampling strategy. Unfreeze all network parameters and perform multi-task joint fine-tuning; S5: In the joint fine-tuning stage, a physical residual loss term based on the conservation law of fluid dynamics is constructed and added to the total loss function; the partial derivatives of the predicted flow field with respect to the spatiotemporal coordinates are calculated using automatic differentiation technology, and the model output is constrained to meet the conditions of mass conservation and momentum conservation. S6: Input the spatiotemporal coordinates to be predicted into the trained neural network model to obtain the cavitation confidence score; map the confidence score to the cavitation probability through the Sigmoid activation function, and reconstruct the continuous liquid phase volume fraction distribution of the entire flow field as the final prediction result of the cavitation flow field.

2. The high-precision prediction method for cavitation flow field of a hydraulic torque converter according to claim 1, characterized in that, The CFD simulation in S1 is based on the Mixture multiphase flow model, in which the cavitation phase change source term is calculated using the Schnerr-Sauer model; the evaporation rate in the Schnerr-Sauer model... With condensation rate The calculation formula is as follows: in, This represents the mass source term generated by liquid vaporization. This represents the mass source term generated by gas condensation; The density of the liquid phase is... The density is the gas phase density. The density of the mixed phase; The preset nucleation volume fraction, Where is the bubble radius; Let be the saturated vapor pressure of the fluid at the current temperature. This represents the local pressure in the flow field.

3. The high-precision prediction method for cavitation flow field of a hydraulic torque converter according to claim 1, characterized in that, In S1, a binary cavitation label is constructed. The specific method is as follows: Let the actual liquid volume fraction obtained from CFD simulation be... The physical threshold for determining cavitation is set as follows: , The value range is from 0.8 to 0.95; for any sample point Its binary emptying label Defined as: in, This indicates that the point is located in a cavitation region. This indicates that the point is located in a non-cavitation region.

4. The high-precision prediction method for cavitation flow field of a hydraulic torque converter according to claim 1, characterized in that, The shared backbone network in S2 is used to extract high-dimensional implicit features of normalized spatiotemporal coordinates; The flow field prediction subnetwork is used to output the velocity field and pressure field based on the implicit features; the cavitation prediction subnetwork is used to output the unnormalized cavitation confidence based on the implicit features. The neural network model employs the Softplus activation function to ensure the continuity of higher-order derivatives when calculating physical residuals in S5. The Softplus activation function is defined as follows: In step S6, the scalar cavitation confidence score output by the cavitation prediction subnetwork is... Mapped to vacuolation probability and the final predicted liquid volume fraction The formula is: in, This represents the Sigmoid function. The range of values ​​is .

5. The high-precision prediction method for cavitation flow field of a hydraulic torque converter according to claim 1, characterized in that, The hybrid supervised loss function in S3 consists of three weighted components: a classification loss term based on the binary cavitation label, a regression correction loss term calculated only in the true cavitation region, and a false positive penalty term calculated only in the true non-cavitation region. The hybrid supervision loss function From classification loss terms Regression Correction Loss Term and false positive penalty items The weighted composition, its mathematical expression is: in, These are the classification loss weight coefficients; These are the regression loss weight coefficients; This represents the false positive penalty weighting coefficient.

6. The high-precision prediction method for cavitation flow field of a hydraulic torque converter according to claim 5, characterized in that, The classification loss item The focus loss function is used, and the calculation formula is as follows: in, To focus parameters, the weights of easily distinguishable samples are reduced so that the model focuses on difficult-to-distinguish samples; This is a balancing parameter used to adjust for the imbalance between positive and negative samples; The probability of the model predicting the true class is defined as: 。 7. The high-precision prediction method for cavitation flow field of a hydraulic torque converter according to claim 5, characterized in that, The regression correction loss term Only applicable to specific cavitation point sets with real labels The calculation uses the mean absolute error form; false positive penalty term. Only for point sets whose real labels are non-empty Calculation; the specific formula is as follows: in, This represents the set of samples from the real cavitation region. This represents the set of samples from the real non-empty regions; It is a linear rectification function, only when the predicted value Below the threshold Punishment will be imposed at that time.

8. The high-precision prediction method for cavitation flow field of a hydraulic torque converter according to claim 5, characterized in that, The physical residual loss term in S5 Including the residuals of the continuity equation for incompressible fluids Residuals of the momentum equation Its definition is as follows: The formula for calculating the total physical residual loss is: in, The velocity vector output by the flow field prediction subnetwork. To predict static pressure, To incorporate the mixed-phase fluid density calculated using cavitation phase change, The dynamic viscosity of the mixed-phase fluid. For the nabla operator, This represents the total number of sample points used to calculate the physical residual.

9. The high-precision prediction method for cavitation flow field of a hydraulic torque converter according to claim 1, characterized in that, The cavitation-aware sampling strategy in S4 specifically involves dividing the training dataset into cavitation sample sets. Non-empty sample set When constructing each training batch, the sampling ratio of cavitation samples is forcibly set to [value missing]. , If the batch size is Then from Randomly selected from One sample, from Randomly selected from A sample, satisfying: in, This indicates rounding a numerical value to the nearest integer; proportion This is used to force the model to acquire sufficient cavitation feature gradients in each iteration step.

10. The high-precision prediction method for cavitation flow field of a hydraulic torque converter according to claim 1, characterized in that, The parameter update logic for the three-stage evolutionary training in S4 is as follows: Let the shared backbone network parameter set be... The parameter set of the flow field prediction subnetwork is The parameter set of the cavitation prediction subnetwork is Phase 1 involves freezing the cavitation prediction subnetwork, pre-training the shared backbone and flow field prediction subnetwork using flow field physical quantity data, and then updating only a subset of parameters in the optimizer. To minimize Phase two involves loading the parameters trained in phase one. The frozen flow field prediction subnetwork utilizes binary cavitation labels and liquid phase volume fraction, and centrally optimizes the hybrid supervised loss function based on a cavitation-sensing sampling strategy. The optimizer only updates a subset of parameters. To minimize ; Phase three involves unfreezing all network parameters, performing joint fine-tuning across multiple tasks, and having the optimizer jointly update all parameters to minimize the total loss. .

Images