A simulation optimization method and system for complex internal flow field of a hydraulic torque converter

Abstract

The application discloses a hydrodynamic torque converter complex internal flow field simulation optimization method and system, relates to the technical field of hydrodynamic torque converters, and comprises the following steps: constructing a hydrodynamic torque converter simulation model and obtaining sample data information of the hydrodynamic torque converter simulation model; training a constructed simulation parameter estimation model through the sample data information; obtaining basic physical quantity information of a hydrodynamic torque converter to be simulated and inputting the basic physical quantity information into the trained simulation parameter estimation model to obtain simulation parameter information, which is used for simulating the hydrodynamic torque converter to be simulated. Simulation data of the constructed hydrodynamic torque converter simulation model is used to form a sample data set, and the constructed simulation parameter estimation model is trained by using the sample data set. The trained model meets the condition of the complex internal flow field of the hydrodynamic torque converter. When any hydrodynamic torque converter needs to be simulated, simulation parameter information can be obtained through known basic physical quantity information, and thus the rapid determination of simulation parameters in the hydrodynamic torque converter flow simulation is realized.

Description

Technical Field

[0001] This application relates to the field of hydraulic torque converter technology, specifically to a simulation optimization method and system for complex internal flow fields in hydraulic torque converters. Background Technology

[0002] HTC (Hydraulic Torque Converter), a core component of modern vehicle transmission systems, achieves flexible torque transmission between the engine and gearbox through fluid kinetic energy. Its structure mainly consists of a pump impeller (input), a turbine (output), and a guide impeller (torque regulation). Compared to mechanical clutches, hydraulic torque converters offer advantages such as automatically adapting to load changes, damping vibrations and shocks, and extending the lifespan of the transmission system. They are widely used in construction machinery, military vehicles, and high-end passenger cars.

[0003] Flow field simulation within a hydraulic torque converter is crucial for verifying its design performance. Traditional design relies on one-dimensional beam theory and empirical formulas, resulting in low accuracy and high experimental costs. Computational fluid dynamics (CFD) simulation technology achieves precise location of energy loss sources through three-dimensional flow field analysis, identifying loss sources such as blade surface separation zones and vortex core locations. CFD simulation of hydraulic torque converters has undergone three stages of development. The first stage was a quasi-steady-state single-channel model, using periodic boundary conditions to simplify calculations and only simulating smooth flow at the design operating point. The second stage was a transient full-channel model, based on sliding interface technology, simulating unsteady interactions between impellers, improving simulation accuracy but also increasing computation time. The third stage involves the integration of deep academic methods, utilizing convolutional neural networks to accelerate smoothness prediction and solve the problem of exploring high-dimensional parameter spaces. With the development of precise models such as turbulence models and multiphase flow simulations, the time required for simulation calculations and parameter tuning has increased dramatically. In particular, the parameter setting and optimization process for simulations requires adjustments to parameters such as geometric defect cleanup, internal flow channel extraction, mesh generation, boundary condition setting, and time step adjustment before simulations can proceed normally. However, this optimization currently relies on engineers' experience, leading to significant differences in simulation success rates and simulation times among different engineers.

[0004] Currently, most researchers focus on optimizing the number of blades in the pump impeller, turbine, and guide vanes of hydraulic torque converters, using intelligent algorithms to optimize the structural dimensions of these components. However, research on the rationality of setting simulation parameters for hydraulic torque converter hydrodynamic simulation has been neglected. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, this application proposes the following technical solution:

[0006] In a first aspect, embodiments of this application provide a method for simulating and optimizing complex internal flow fields in a hydraulic torque converter, including:

[0007] Construct a hydraulic torque converter simulation model and obtain sample data information of the hydraulic torque converter simulation model;

[0008] The simulation parameter estimation model is trained using the sample data information.

[0009] The basic physical quantity information of the hydraulic torque converter to be simulated is obtained and input into the trained simulation parameter estimation model to obtain simulation parameter information, which is used to simulate the hydraulic torque converter to be simulated.

[0010] In one possible implementation, constructing a general hydraulic torque converter simulation model applicable to hydraulic torque converters includes:

[0011] Geometric modeling and processing were performed to determine the three-dimensional model of the hydraulic torque converter, clean up geometric defects, and extract its internal flow channels.

[0012] Finite element mesh generation: Define the boundary layer mesh, impeller flow channel structured mesh, and shell region unstructured mesh for the hydraulic torque converter model;

[0013] Record the set values ​​for the boundary layer first layer height, expansion ratio, impeller channel mesh size, and total mesh quantity parameters;

[0014] Physical model settings: Based on the hydraulic torque converter, set the rotational domain rating and turbulence model, and record the speed setting and model selection.

[0015] Boundary condition setting: Set and record the inlet and outlet settings of the hydraulic torque converter;

[0016] Solver configuration: Set and record the solution strategy selection and convergence control parameter settings.

[0017] In one possible implementation, the step of constructing a hydraulic torque converter simulation model and obtaining sample data information of the hydraulic torque converter simulation model includes:

[0018] Construct a general-purpose hydraulic torque converter simulation model applicable to hydraulic torque converters;

[0019] Multiple simulations were performed using the hydraulic torque converter simulation model, and flow field slices were extracted from the simulation results to form a sample dataset.

[0020] In one possible implementation, the step of performing multiple simulations using the hydraulic torque converter simulation model and extracting flow field slices from the simulation results to form a sample dataset includes:

[0021] The RGB three channels of the flow field slice are determined, and the vector field representing the velocity is used to obtain the velocity cloud map in the simulation model of the hydraulic torque converter.

[0022] Based on Q-criterion preprocessing, the Galilean invariant Q>0 is used to represent the vortex structure to determine the vortex diagram in the simulation model of the hydraulic torque converter. The expression is:

[0023]

[0024] In the formula: Let Frobenius norm be the matrix. A Let be the velocity gradient tensor of the symmetric part. B The velocity gradient tensor for the antisymmetric part.

[0025] In one possible implementation, training the constructed simulation parameter estimation model using the sample data information includes:

[0026] The sample data in the sample dataset is normalized to obtain the normalized sample dataset.

[0027] The normalized sample dataset is divided into a training set, a test set, and a validation set;

[0028] Initialize the simulation parameter estimation model and multi-task loss function, set the network learning rate, and train the network using the Adam backpropagation optimization algorithm;

[0029] Traverse the training set sample batches, perform network forward propagation calculations, calculate the loss function error and gradient based on the network output and label values, and update the neuron weights and biases.

[0030] Repeat the training until all data in the training set has been calculated, use the validation set for evaluation, record the training time for this round, and adjust the learning rate based on the results;

[0031] Training continues until the model reaches the convergence condition set by the model, then the simulation parameter estimation model parameter file is saved, and the model training ends.

[0032] In one possible implementation, the simulation parameter estimation model includes a multi-scale feature extraction module, a cross-modal fusion module, and a eddy detection and evaluation module. The multi-scale feature extraction module receives the velocity contour map and eddy current map and extracts feature information from them. The cross-modal fusion module performs multi-scale feature fusion of the feature information, fusing feature maps of different resolutions through bilinear pooling to improve the representation ability and robustness of the features. The eddy detection and evaluation module uses the feature map output by the cross-modal fusion module to determine whether there are eddies and evaluate the eddy mass.

[0033] In one possible implementation, the multi-scale feature extraction module adopts a heterogeneous dual-branch architecture, including: a velocity feature extraction branch and a vorticity feature extraction branch;

[0034] The velocity feature extraction branch is based on an improved ResNet-50, using only the first three residual blocks. A vorticity-guided CBAM attention mechanism is embedded within these residual blocks. Channel attention dynamically focuses on high vorticity regions, while spatial attention sharpens the vortex kernel boundaries. The channel attention weight calculation formula is as follows:

[0035]

[0036]

[0037] in: For activation function, MPL It is a multilayer perceptron. GAP For global average pooling, F This is a velocity field characteristic map. C The number of feature channels, H × W The feature map size;

[0038] The vortex feature extraction branch employs an improved U-Net encoder-decoder. The encoder gradually reduces the spatial resolution of the feature map through multiple convolution and pooling operations, effectively extracting the contextual information of the image and gradually capturing a larger receptive field. The decoder gradually restores the spatial resolution of the feature map through upsampling operations, and at the same time combines skip connections to concatenate the high-resolution feature map in the encoder with the upsampled feature map, preserving detailed information and enhancing feature fusion.

[0039] In one possible implementation, the cross-modal fusion module employs a weighted bidirectional feature pyramid network (BiFPN) to extract features, which is divided into two parts: bottom-up feature transfer and top-down feature transfer.

[0040] Bottom-up feature propagation: BiFPN starts from the bottom layer, uses bilinear pooling to upsample low-resolution feature maps to high resolution, and then uses bidirectional connections to fuse the feature maps of the previous layer with the upsampled feature maps of the next layer.

[0041] Top-down feature propagation: Next, BiFPN propagates features along the top-down path of the feature pyramid network. Using bidirectional connections, it fuses the feature maps of the previous layer with the upsampled feature maps of the next layer to obtain richer and more accurate feature representations.

[0042] In one possible implementation, the total loss function for training the simulation parameter estimation model is:

[0043]

[0044] in, To converge the weights of the classification task, For time regression task weights, Did the simulation successfully classify the loss? To determine the classification loss for simulation convergence; This represents the simulation time regression loss.

[0045] The for:

[0046] ;

[0047] Where: N is the number of data samples. Let the label indicate whether the simulation of the i-th sample was successful. Let be the simulation success probability predicted by the model for the i-th sample. These are the weighting coefficients for positive samples;

[0048] for:

[0049]

[0050] in: This is the index set of successful simulation samples. The number of successful simulation samples. This is the label for whether the simulation converged for the i-th sample, with 1 for success and 0 for failure. Let be the simulation convergence probability predicted by the model for the i-th sample.

[0051] for:

[0052] ;

[0053] Among them: Among them, For the simulation convergence sample index set, To simulate the number of convergent samples, For the actual convergence time, To predict the convergence time, For safety and convenience, threshold values, This is the boundary penalty coefficient.

[0054] Secondly, embodiments of this application provide a simulation and optimization system for complex internal flow fields of a hydraulic torque converter, including:

[0055] The sample data acquisition module is used to construct a hydraulic torque converter simulation model and acquire sample data information of the hydraulic torque converter simulation model;

[0056] The model training module is used to train the constructed simulation parameter estimation model using the sample data information;

[0057] The simulation parameter acquisition module is used to acquire the basic physical quantity information of the hydraulic torque converter to be simulated and input it into the trained simulation parameter estimation model to obtain simulation parameter information. The simulation parameter information is used to simulate the hydraulic torque converter to be simulated.

[0058] In this embodiment, a sample dataset is formed by constructing a hydraulic torque converter simulation model, and the constructed simulation parameter estimation model is trained using this sample dataset. The trained model thus satisfies the conditions for the complex internal flow field of the hydraulic torque converter. When simulating any hydraulic torque converter, the simulation parameter information can be obtained using known basic physical quantity information, thereby achieving rapid determination of simulation parameters during hydraulic torque converter flow simulation. Attached Figure Description

[0059] Figure 1 A flowchart illustrating a method for simulating and optimizing a complex internal flow field in a hydraulic torque converter, as provided in this application embodiment;

[0060] Figure 2 This is a schematic diagram of the structure of the simulation parameter estimation model provided in the embodiments of this application;

[0061] Figure 3 This is a schematic diagram of the structure of BiFPN provided in an embodiment of this application;

[0062] Figure 4 This is a schematic diagram of a complex internal flow field simulation and optimization system for a hydraulic torque converter, provided as an embodiment of this application. Detailed Implementation

[0063] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0064] See Figure 1 The simulation and optimization method for complex internal flow fields of a hydraulic torque converter provided in this embodiment includes:

[0065] S101, Construct a hydraulic torque converter simulation model and obtain sample data information of the hydraulic torque converter simulation model.

[0066] This embodiment first constructs a simulation model of a hydraulic torque converter, mainly including the following steps: Geometric modeling and processing: determining the three-dimensional model of the hydraulic torque converter, cleaning up geometric defects, and extracting its internal flow channels. Finite element mesh generation: defining the boundary layer mesh, the structured mesh of the impeller flow channel, and the unstructured mesh of the shell region of the hydraulic torque converter model. Recording parameter settings such as the first layer height of the boundary layer, expansion ratio, impeller flow channel mesh size, and total mesh quantity. Physical model setting: setting the rated values ​​of the rotating domain and the turbulence model according to the hydraulic torque converter, and recording the speed setting values ​​and model selection. Boundary condition setting: setting and recording the inlet and outlet settings of the hydraulic torque converter. Solver configuration: setting and recording the solution strategy selection and convergence control parameter settings.

[0067] After the hydraulic torque converter simulation model is constructed, multiple simulations are performed using the hydraulic torque converter simulation model, and flow field slices are extracted from the simulation results to form a sample dataset.

[0068] Specifically, the following data is generated by extracting flow field slices from the simulation results:

[0069] (1) Velocity contour plot. The vector field of velocity is represented by three RGB channels. The R (red) channel is the normalized value of the X-direction component (u), the G (green) channel is the normalized value of the Y-direction component (v), and the B (blue) channel is the normalized value of the Z-direction component (w).

[0070] (2) Vorticity diagram, single channel, preprocessed based on the Q criterion, using the Galilean invariant Q>0 to represent the vortex structure, the expression is:

[0071]

[0072] In the formula, Let Frobenius norm be the matrix. A Let be the velocity gradient tensor of the symmetric part. B The velocity gradient tensor for the antisymmetric part.

[0073] S102, The constructed simulation parameter estimation model is trained using the sample data information.

[0074] In this embodiment, the training process of the simulation parameter estimation model is as follows: The sample data in the sample dataset is normalized to obtain a normalized sample dataset. The normalized sample dataset is divided into a training set, a test set, and a validation set. The simulation parameter estimation model and multi-task loss function are initialized, the network learning rate is set, and the Adam backpropagation optimization algorithm is used for training. The training set sample batches are traversed, and forward propagation calculations are performed. The loss function error and gradient are calculated based on the network output and label values, and the neuron weights and biases are updated. Training is repeated until all data in the training set has been calculated. The validation set is used for evaluation, the training time for this round is recorded, and the learning rate is adjusted based on the results until the training reaches the convergence condition set by the model. The simulation parameter estimation model parameter file is saved, and the model training ends.

[0075] See Figure 2 The simulation parameter estimation model network structure constructed in this embodiment is divided into the following parts: multi-scale feature extraction module, cross-modal fusion module, and eddy detection and evaluation module.

[0076] The multi-scale feature extraction module adopts a heterogeneous dual-branch architecture: the velocity feature extraction branch is based on the improved ResNet-50, using only the first 3 residual blocks, embedding a vorticity-guided CBAM attention mechanism in the residual blocks, dynamically focusing on high vorticity regions through channel attention, and sharpening the vortex kernel boundary through spatial attention.

[0077] The formula for calculating channel attention weights is as follows:

[0078]

[0079]

[0080] in: For activation function, MPL It is a multilayer perceptron. GAP For global average pooling, F This is a velocity field characteristic map. C The number of feature channels, H × W The size of the feature map.

[0081] Taking a velocity cloud map with an input size of 512×512×3 as an example, the network structure and input / output of each layer of the velocity feature extraction branch are shown in Table 1 below. Finally, the output feature maps of convolutional blocks conv block2 and conv block4 are input into the cross-modal fusion module.

[0082] Table 1. Network structure and input / output of each layer in the velocity feature extraction branch.

[0083]

[0084] The vortex feature extraction branch employs an improved U-Net encoder-decoder. The encoder gradually reduces the spatial resolution of the feature map through multiple convolution and pooling operations, effectively extracting the contextual information of the image and gradually capturing a larger receptive field. The decoder gradually restores the spatial resolution of the feature map through upsampling operations, and at the same time combines skip connections to concatenate the high-resolution feature map in the encoder with the upsampled feature map, preserving detailed information and enhancing feature fusion.

[0085] Taking a velocity contour map with an input size of 512×512×1 as an example, the network structure and input / output of each layer of the vorticity feature extraction branch are shown in Table 2 below. Finally, the output feature maps of the convolutional blocks decode block2 and decode block3 are input into the cross-modal fusion module.

[0086] Table 2. Network structure and input / output of each layer in the vorticity feature extraction branch.

[0087]

[0088] The cross-modal fusion module uses a weighted bidirectional feature pyramid network (BiFPN) to extract features, and consists of two parts: bottom-up feature propagation and top-down feature propagation.

[0089] Bottom-up feature propagation: BiFPN starts from the bottom layer, using bilinear pooling to upsample low-resolution feature maps to high resolution, and then uses bidirectional connections to fuse the feature maps from the previous layer with the upsampled feature maps from the next layer. This bottom-up feature propagation helps to obtain richer information from lower levels.

[0090] Top-down feature propagation: Next, BiFPN propagates features along the top-down path of the feature pyramid network. In this process, BiFPN utilizes bidirectional connections to fuse the feature maps of the previous layer with the upsampled feature maps of the next layer to obtain richer and more accurate feature representations.

[0091] BiFPN overall structure diagram Figure 3 As shown, multi-scale feature fusion is performed at each level, and feature maps of different resolutions are fused through bilinear pooling, thereby improving the representation ability and robustness of features. P 3 P 4 P 5 P 6 P The outputs corresponding to 7 are as follows:

[0092]

[0093]

[0094]

[0095]

[0096]

[0097] The feature map output by the cross-modal fusion module is passed through a fully connected layer to determine whether there are vortices and to evaluate the quality of the vortices.

[0098] In this embodiment, the total loss function for training the simulation parameter estimation model is:

[0099]

[0100] in, To converge the weights of the classification task, For time regression task weights, Did the simulation successfully classify the loss? To determine the classification loss for simulation convergence; This represents the simulation time regression loss.

[0101] Specifically, the for:

[0102] ;

[0103] Where: N is the number of data samples. Let the label indicate whether the simulation of the i-th sample was successful. Let be the simulation success probability predicted by the model for the i-th sample. These are the weighting coefficients for positive samples;

[0104] for:

[0105]

[0106] in: This is the index set of successful simulation samples. The number of successful simulation samples. This is the label for whether the simulation converged for the i-th sample, with 1 for success and 0 for failure. Let be the simulation convergence probability predicted by the model for the i-th sample.

[0107] for:

[0108] ;

[0109] Among them: Among them, For the simulation convergence sample index set, To simulate the number of convergent samples, For the actual convergence time, To predict the convergence time, For safety and convenience, threshold values, This is the boundary penalty coefficient.

[0110] S103: Obtain the basic physical quantity information of the hydraulic torque converter to be simulated and input it into the trained simulation parameter estimation model to obtain simulation parameter information. The simulation parameter information is used to simulate the hydraulic torque converter to be simulated.

[0111] Corresponding to the simulation and optimization method for complex internal flow field of a hydraulic torque converter provided in the above embodiments, this application also provides an embodiment of a simulation and optimization system for complex internal flow field of a hydraulic torque converter.

[0112] See Figure 4 A simulation and optimization system for complex internal flow fields in hydraulic torque converters, comprising:

[0113] The sample data acquisition module 201 is used to construct a hydraulic torque converter simulation model and acquire sample data information of the hydraulic torque converter simulation model.

[0114] The model training module 202 is used to train the constructed simulation parameter estimation model using the sample data information.

[0115] The simulation parameter acquisition module 203 is used to acquire the basic physical quantity information of the hydraulic torque converter to be simulated and input it into the trained simulation parameter estimation model to obtain simulation parameter information. The simulation parameter information is used to simulate the hydraulic torque converter to be simulated.

[0116] The velocity cloud map and vorticity map are used to determine the boundary layer height, expansion ratio, impeller channel mesh size, total mesh size, rotational speed, turbulence model type, hydraulic torque converter inlet setting, hydraulic torque converter outlet setting, solution strategy setting, and convergence control parameter setting in the hydraulic torque converter simulation model.

[0117] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0118] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A simulation and optimization method for complex internal flow fields in a hydraulic torque converter, characterized in that, include: Constructing a hydraulic torque converter simulation model and obtaining sample data information of the hydraulic torque converter simulation model, including: Construct a general-purpose hydraulic torque converter simulation model applicable to hydraulic torque converters; Multiple simulations were performed using the hydraulic torque converter simulation model, and flow field slices were extracted from the simulation results to form a sample dataset. This included: determining the RGB three channels of the flow field slices, using the RGB three channels to represent the velocity vector field, and obtaining the velocity cloud map in the hydraulic torque converter simulation model; based on Q-criterion preprocessing, using the Galilean invariant Q>0 to represent the vortex structure to determine the vortex volume diagram in the hydraulic torque converter simulation model. The constructed simulation parameter estimation model is trained using the sample data information, wherein: The simulation parameter estimation model includes a multi-scale feature extraction module, a cross-modal fusion module, and a eddy detection and evaluation module. The multi-scale feature extraction module receives the velocity contour map and eddy current map, and extracts feature information from them. The cross-modal fusion module performs multi-scale feature fusion of the feature information, fusing feature maps of different resolutions through bilinear pooling, thereby improving the representation ability and robustness of the features. The eddy detection and evaluation module uses the feature map output by the cross-modal fusion module to determine whether there are eddies and evaluates the eddy mass. The multi-scale feature extraction module adopts a heterogeneous dual-branch architecture, including: a velocity feature extraction branch and a vorticity feature extraction branch; the velocity feature extraction branch is based on an improved ResNet-50, using only the first 3 residual blocks, embedding a vorticity-guided CBAM attention mechanism in the residual blocks, dynamically focusing on high vorticity regions through channel attention, and sharpening the vortex kernel boundary through spatial attention. The vorticity feature extraction branch employs an improved U-Net encoder-decoder. The encoder gradually reduces the spatial resolution of the feature map through multiple convolution and pooling operations, effectively extracting the contextual information of the image and gradually capturing a larger receptive field. The decoder gradually restores the spatial resolution of the feature map through upsampling operations, and at the same time combines skip connections to concatenate the high-resolution feature map in the encoder with the upsampled feature map, preserving detailed information and enhancing feature fusion. The cross-modal fusion module uses a weighted bidirectional feature pyramid network (BiFPN) to extract features, and is divided into two parts: bottom-up feature transfer and top-down feature transfer. The basic physical quantity information of the hydraulic torque converter to be simulated is obtained and input into the trained simulation parameter estimation model to obtain simulation parameter information, which is used to simulate the hydraulic torque converter to be simulated.

2. The method for simulating and optimizing the complex internal flow field of a hydraulic torque converter according to claim 1, characterized in that, The construction of a general hydraulic torque converter simulation model applicable to hydraulic torque converters includes: Geometric modeling and processing were performed to determine the three-dimensional model of the hydraulic torque converter, clean up geometric defects, and extract its internal flow channels. Finite element mesh generation: Define the boundary layer mesh, the structured mesh of the impeller flow channel, and the unstructured mesh of the shell region for the hydraulic torque converter model; Record the set values ​​for the boundary layer first layer height, expansion ratio, impeller channel mesh size, and total mesh quantity parameters; Physical model settings: Based on the hydraulic torque converter, set the rotational domain rating and turbulence model, and record the speed setting and model selection. Boundary condition setting: Set and record the inlet and outlet settings of the hydraulic torque converter; Solver configuration: Set and record the solution strategy selection and convergence control parameter settings.

3. The method for simulating and optimizing the complex internal flow field of a hydraulic torque converter according to claim 1, characterized in that, The expression for the Galilean invariant Q is: In the formula: Let Frobenius norm be the matrix. A Let be the velocity gradient tensor of the symmetric part. B The velocity gradient tensor for the antisymmetric part.

4. The method for simulating and optimizing the complex internal flow field of a hydraulic torque converter according to claim 1, characterized in that, The step of training the constructed simulation parameter estimation model using the sample data information includes: The sample data in the sample dataset is normalized to obtain the normalized sample dataset. The normalized sample dataset is divided into a training set, a test set, and a validation set; Initialize the simulation parameter estimation model and multi-task loss function, set the network learning rate, and train it using the Adam backpropagation optimization algorithm; Traverse the training set sample batches, perform network forward propagation calculations, calculate the loss function error and gradient based on the network output and label values, and update the neuron weights and biases. Repeat the training until all data in the training set has been calculated, use the validation set for evaluation, record the training time for this round, and adjust the learning rate based on the results; Training continues until the model reaches the convergence condition set by the model, then the simulation parameter estimation model parameter file is saved, and the model training ends.

5. The method for simulating and optimizing the complex internal flow field of a hydraulic torque converter according to claim 1, characterized in that, The formula for calculating channel attention weights is: in: For activation function, MPL It is a multilayer perceptron. GAP For global average pooling, F This is a velocity field characteristic map. C The number of feature channels, H × W The size of the feature map.

6. The method for simulating and optimizing the complex internal flow field of a hydraulic torque converter according to claim 1, characterized in that, Bottom-up feature propagation: BiFPN starts from the bottom layer, uses bilinear pooling to upsample low-resolution feature maps to high resolution, and then uses bidirectional connections to fuse the feature maps of the previous layer with the upsampled feature maps of the next layer. Top-down feature propagation: Next, BiFPN propagates features along the top-down path of the feature pyramid network. Using bidirectional connections, it fuses the feature maps of the previous layer with the upsampled feature maps of the next layer to obtain richer and more accurate feature representations.

7. The method for simulating and optimizing the complex internal flow field of a hydraulic torque converter according to claim 1, characterized in that, The total loss function for training the simulation parameter estimation model is: in, To converge the weights of the classification task, For time regression task weights, Did the simulation successfully classify the loss? To determine the classification loss for simulation convergence; For simulation time regression loss; The for: ； Where: N is the number of data samples. Let the i-th sample be labeled as indicating whether the simulation was successful. Let be the simulation success probability predicted by the model for the i-th sample. These are the weighting coefficients for positive samples; for: in: This is the index set of successful simulation samples. The number of successful simulation samples. This is the label for whether the simulation converged for the i-th sample, with 1 for success and 0 for failure. Let be the simulation convergence probability predicted by the model for the i-th sample. for: ； Among them: Among them, For the simulation convergence sample index set, To simulate the number of convergent samples, For the actual convergence time, To predict the convergence time, For safety and convenience, threshold values, This is the boundary penalty coefficient.

8. A simulation and optimization system for complex internal flow fields in a hydraulic torque converter, characterized in that, include: The sample data acquisition module is used to construct a hydraulic torque converter simulation model and acquire sample data information of the hydraulic torque converter simulation model, including: Construct a general-purpose hydraulic torque converter simulation model applicable to hydraulic torque converters; Multiple simulations were performed using the hydraulic torque converter simulation model, and flow field slices were extracted from the simulation results to form a sample dataset. This included: determining the RGB three channels of the flow field slices, using the RGB three channels to represent the velocity vector field, and obtaining the velocity cloud map in the hydraulic torque converter simulation model; based on Q-criterion preprocessing, using the Galilean invariant Q>0 to represent the vortex structure to determine the vortex volume diagram in the hydraulic torque converter simulation model. The model training module is used to train the constructed simulation parameter estimation model using the sample data information, wherein: The simulation parameter estimation model includes a multi-scale feature extraction module, a cross-modal fusion module, and a eddy detection and evaluation module. The multi-scale feature extraction module receives the velocity contour map and eddy current map, and extracts feature information from them. The cross-modal fusion module performs multi-scale feature fusion of the feature information, fusing feature maps of different resolutions through bilinear pooling, thereby improving the representation ability and robustness of the features. The eddy detection and evaluation module uses the feature map output by the cross-modal fusion module to determine whether there are eddies and evaluates the eddy mass. The multi-scale feature extraction module adopts a heterogeneous dual-branch architecture, including: a velocity feature extraction branch and a vorticity feature extraction branch; the velocity feature extraction branch is based on an improved ResNet-50, using only the first 3 residual blocks, embedding a vorticity-guided CBAM attention mechanism in the residual blocks, dynamically focusing on high vorticity regions through channel attention, and sharpening the vortex kernel boundary through spatial attention. The vorticity feature extraction branch employs an improved U-Net encoder-decoder. The encoder gradually reduces the spatial resolution of the feature map through multiple convolution and pooling operations, effectively extracting the contextual information of the image and gradually capturing a larger receptive field. The decoder gradually restores the spatial resolution of the feature map through upsampling operations, and at the same time combines skip connections to concatenate the high-resolution feature map in the encoder with the upsampled feature map, preserving detailed information and enhancing feature fusion. The cross-modal fusion module uses a weighted bidirectional feature pyramid network (BiFPN) to extract features, and is divided into two parts: bottom-up feature transfer and top-down feature transfer. The simulation parameter acquisition module is used to acquire the basic physical quantity information of the hydraulic torque converter to be simulated and input it into the trained simulation parameter estimation model to obtain simulation parameter information. The simulation parameter information is used to simulate the hydraulic torque converter to be simulated.

Images