A priori guided residual prediction method for fluid torque converter cascade flow field

Abstract

This invention relates to the interdisciplinary field of flow field simulation and prediction for hydraulic components and deep learning. It provides a priori-guided residual prediction method for the flow field of a hydraulic torque converter cascade. First, it acquires the full-channel flow field data of the target cascade under multiple known operating conditions and constructs an inlet condition function. Then, it applies the inlet condition function to the single-channel computational domain of the target cascade to obtain multi-condition single-channel simulation data, training a single-channel multi-condition prior prediction model. Subsequently, it uses the single-channel multi-condition prior prediction model to obtain the single-channel predicted flow field under the predicted operating condition, and maps it to the full-channel of the target cascade according to the period angle of the target cascade, obtaining a directly mapped prior flow field. Finally, it compensates using a residual correction model to obtain the corrected predicted pressure and velocity fields of the full-channel of the target cascade. This invention can reduce the cost of repeated calculations of the full-channel of the target cascade under multiple operating conditions and improve the rapid prediction efficiency of the flow field of the hydraulic torque converter cascade.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of flow field simulation and prediction of hydraulic components and deep learning, and particularly relates to a priori guided residual prediction method for the flow field of a hydraulic torque converter blade cascade. Background Technology

[0002] A hydraulic torque converter consists of a pump impeller, a turbine, and a guide impeller. The pressure and velocity field distribution within its circumferentially periodic flow channel structure's blade cascade components are closely related to energy transfer efficiency, flow losses, and operational stability. Rapid prediction of the flow field of a target blade cascade is a crucial foundation for blade structure design, performance analysis, and optimization. The target blade cascade's full flow channel, as described in this invention, refers to the fluid region formed by all flow channels corresponding to the target blade cascade, and is not limited to the entire flow channel of the entire machine formed by the integral coupling of the pump impeller, turbine, and guide impeller.

[0003] Currently, the flow field of hydraulic torque converter blades is usually analyzed using computational fluid dynamics methods. Full-channel simulation of the target blade cascade can relatively completely reflect the periodic structure, circumferential non-uniformity, and inlet boundary characteristics of the blade cascade; however, the large mesh size and long computation time make it difficult to meet the needs of multi-condition prediction and rapid optimization.

[0004] To reduce computational load, single-channel or periodic channel models are often used in engineering for local flow simulation. However, the results of single-channel models are significantly affected by inlet boundary conditions, and when extended to the entire channel, pressure scalar mapping and velocity vector direction transformation need to be handled. Therefore, the results cannot be directly equated to the flow field of the entire target cascade and are more suitable as prior information for the prediction of the entire channel.

[0005] With the development of data-driven modeling methods, establishing the mapping relationship between operating conditions and flow field response using surrogate models has become an important way to improve the efficiency of flow field prediction. However, existing surrogate models usually only use scalar operating parameters such as speed ratio and flow coefficient as input, which makes it difficult to fully characterize the influence of inlet velocity magnitude, direction, and radial distribution on the single-channel flow field as the operating conditions change, thus limiting the model's prediction ability under intermediate operating conditions or operating conditions not used in training. In addition, although there are schemes that apply residual learning to flow field prediction, they are mainly aimed at general fluid simulation scenarios and have not been systematically designed for the circumferential periodicity of hydraulic torque converter blades and the special requirements of single-channel to full-channel mapping.

[0006] In summary, there is currently a lack of a flow field prediction method for cascades that can balance the computational accuracy of the entire flow channel of the target cascade with the need for rapid prediction under multiple operating conditions, while making full use of inlet boundary information and effectively compensating for periodic mapping errors. Summary of the Invention

[0007] The purpose of this invention is to provide a priori guided residual prediction method for the flow field of a hydraulic torque converter blade cascade, aiming to solve the problems mentioned in the background art.

[0008] The present invention is implemented as follows: a priori guided residual prediction method for the flow field of a hydraulic torque converter blade cascade includes the following steps:

[0009] Step 1: Obtain the full-channel flow field data of the target blade cascade under multiple known operating conditions; the full-channel flow field data includes the spatial coordinates, pressure field, velocity field, and inlet cross-section velocity distribution within the full-channel of the target blade cascade.

[0010] Step 2: Based on the known velocity distribution of the inlet section of the target blade cascade under the known operating conditions, the inlet velocity is decomposed into components, subjected to radial statistics, interpolation or fitting to obtain the inlet condition function that characterizes the changes in the inlet boundary under different operating conditions; the inlet condition function is used to describe the magnitude, direction and spatial distribution of the target blade cascade inlet velocity as a function of the operating conditions.

[0011] Step 3: Construct a single-channel computational domain based on the circumferential periodicity of the target cascade, and apply the inlet condition functions corresponding to different operating conditions as inlet boundary conditions to the single-channel computational domain to perform single-channel flow field simulation and obtain single-channel simulation data under multiple operating conditions; train a single-channel multi-operating-condition prior prediction model with inlet function conditionalization based on the single-channel simulation data; the single-channel multi-operating-condition prior prediction model takes the single-channel spatial coordinates and inlet condition functions as inputs, and the single-channel pressure field and velocity field as outputs;

[0012] Step 4: Use the trained single-channel multi-condition prior prediction model to obtain the single-channel predicted flow field under known or predicted conditions. Based on the circumferential position and rotation angle of the target blade cascade, periodically map the single-channel predicted flow field to the entire flow channel of the target blade cascade to obtain the directly mapped prior flow field.

[0013] Step 5: Construct a residual correction model, using the direct-mapped prior flow field, the spatial coordinates of the target cascade's entire flow channel, and the inlet condition function as inputs, and the pressure and velocity residuals as outputs, to learn the residuals of the direct-mapped prior flow field relative to the reference flow field of the target cascade's entire flow channel. In the prediction phase, the direct-mapped prior flow field corresponding to the operating condition to be predicted is input into the trained residual correction model to obtain the prediction residuals. The prediction residuals are then superimposed with the direct-mapped prior flow field to obtain the corrected prediction result of the target cascade's entire flow channel.

[0014] A further technical solution is that the target blade cascade is a blade cascade component with a circumferential periodic flow channel structure in a hydraulic torque converter, including a guide wheel blade cascade, a pump wheel blade cascade, or a turbine blade cascade.

[0015] The target blade cascade full flow channel refers to all flow channel regions corresponding to the target blade cascade, and the single flow channel calculation domain is a local flow channel calculation domain selected from the target blade cascade full flow channel according to the circumferential periodicity of the target blade cascade.

[0016] A further technical solution, in step 2, includes the following method for constructing the entry condition function:

[0017] The inlet velocity sampling points and their spatial coordinates are extracted from the inlet section of the entire flow channel of the target cascade. The inlet velocity sampling points include velocity vectors at different radial and circumferential positions on the inlet section.

[0018] Using the rotation axis of the hydraulic torque converter as the axial direction, a cylindrical coordinate description of the inlet section is established. For any velocity sampling point on the inlet section, the radial and circumferential positions are determined according to its spatial coordinates, and the velocity vector of that point is decomposed into axial velocity components, tangential velocity components, and radial velocity components.

[0019] The inlet section is divided into several radial intervals along the radial direction of the inlet section for radial binning statistics. Velocity sampling points whose radial positions fall within the same radial interval are grouped into the same bin. The velocity components at different circumferential positions in each bin are averaged or weighted averaged to obtain the representative values ​​of axial velocity, tangential velocity and radial velocity corresponding to the radial interval. Thus, the two-dimensional velocity distribution on the inlet section is transformed into a three-component inlet velocity profile that varies with the radial position.

[0020] Interpolation or fitting is performed on the three-component inlet velocity profile obtained from multiple radial bins to obtain the inlet condition function used to characterize the changes in the inlet boundary under different working conditions.

[0021] A further technical solution is that, in step 2, for the working condition to be predicted located between two adjacent known working conditions, the entry condition function corresponding to the working condition to be predicted is generated in the following way:

[0022] First, obtain the inlet condition functions corresponding to the two adjacent known working conditions; then, for each radial position, perform interpolation calculations on the axial velocity component, tangential velocity component and radial velocity component in the working direction to obtain the velocity component functions under the working condition to be predicted, thereby forming the inlet condition function corresponding to the working condition to be predicted.

[0023] A further technical solution is that the single-channel multi-condition prior prediction model includes a spatial coding module, an inlet condition coding module, a feature fusion module, and a flow field decoding module;

[0024] The spatial coding module takes single-channel spatial coordinates, radial position features and boundary labels as input, and uses multi-scale position feature coding and a two-layer fully connected network to extract spatial latent features.

[0025] The entry condition coding module takes the operating parameters, entry reference speed and entry condition function coding vector as input, and uses low-frequency condition feature coding and a two-layer fully connected network to extract the latent features of the entry condition.

[0026] The feature fusion module is used to concatenate spatial latent features and entry condition latent features to obtain fused features;

[0027] The flow field decoding module is used to output the pressure component and three-dimensional velocity component at the node based on the fusion features.

[0028] In a further technical solution, in step 4, the periodic mapping adopts a reverse mapping method:

[0029] Identify the circumferential position and rotation angle of the target blade cascade's full-channel node, and reverse map the full-channel node to the reference single-channel coordinate system according to the corresponding period angle. Input the reverse-mapped single-channel coordinate features and the inlet condition function of the corresponding working condition into the trained single-channel multi-working-condition prior prediction model to obtain the pressure prior value and velocity prior value of the node in the reference single-channel coordinate system. Map the pressure prior value and velocity prior value back to the target blade cascade's full-channel coordinate system, where the pressure remains unchanged as a scalar and the velocity is transformed by directional rotation according to the corresponding period angle.

[0030] In a further technical solution, in step 5, the input of the residual correction model also includes the pressure field and velocity field in the direct mapping prior flow field; when training the residual correction model, the parameters of the single-channel multi-condition prior prediction model are kept unchanged, so that the residual correction model learns the systematic deviation between the direct mapping prior flow field and the reference flow field of the target blade cascade in the entire flow channel.

[0031] The prior-guided residual prediction method for the flow field of a hydraulic torque converter blade cascade provided in this embodiment of the invention has the following beneficial effects.

[0032] (1) This method extracts the inlet cross-sectional velocity distribution from the full flow field of the target cascade under known operating conditions and constructs the inlet condition function, enabling multi-condition flow field prediction to utilize inlet boundary information that is closer to the actual flow state. Compared with using only the speed ratio, flow coefficient or other scalar operating condition parameters as model input, it can more fully characterize the influence of inlet velocity magnitude, direction and radial distribution on the changes of operating conditions.

[0033] (2) Single-channel simulation is performed using inlet condition functions, and a single-channel multi-condition prior prediction model with inlet function conditionation is trained based on the single-channel simulation data. This method can obtain the single-channel flow field response under different conditions at a lower computational cost, providing prior information for the prediction of the full-channel flow field of the target cascade, and avoiding the need to carry out high-cost full-channel CFD simulation of the target cascade for each condition to be predicted.

[0034] (3) The predicted flow field of a single flow channel is periodically mapped according to the period angle of the target blade cascade, and the pressure scalar mapping and velocity vector direction rotation transformation are processed separately. This method can improve the coordinate consistency and physical quantity expression accuracy when the prediction results of a single flow channel are extended to the entire flow channel of the target blade cascade.

[0035] (4) A residual correction stage is set up based on the directly mapped prior flow field to compensate for the residual deviation between the directly mapped prior flow field and the reference flow field of the target cascade's entire flow channel. The residual deviation originates from the interpolation or fitting error of the inlet condition function, the single-channel prior prediction error, the full-channel periodic mapping error, and the velocity vector rotation transformation error, and has a clear physical and mapping source. Unlike the method of directly predicting the full-channel flow field of the target cascade from spatial coordinates and operating parameters, this invention uses the directly mapped prior flow field formed by the single-channel prior prediction result as the basis for residual correction, so that the residual correction model mainly learns the systematic deviations in the directly mapped prior flow field that have not been predicted by the single channel and expressed by the periodic mapping, thereby reducing the difficulty of predicting the full-channel flow field and improving the consistency between the correction result and the reference flow field.

[0036] (5) This method can be applied to the prediction of the pressure field and velocity field of the target blade cascade under multiple known working conditions and the working conditions to be predicted. It can be used for the design of hydraulic torque converter blade structure, working condition analysis and rapid performance evaluation. Attached Figure Description

[0037] Figure 1 A flowchart of a priori guided residual prediction method for the flow field of a hydraulic torque converter blade cascade provided in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the entire flow channel, inlet section, and single-channel computational domain of the target cascade. Figure 2 'a' represents the full flow path model of the guide vane cascade. Figure 2 b represents the single-channel model of the guide vane cascade. Figure 2 c represents the inlet section of the entire flow channel of the guide vane. Figure 2 d represents the inlet section of a single flow channel in the guide vane cascade;

[0039] Figure 3 The resulting graph shows the inlet condition function under the OP=0.35 operating condition. Figure 3 'a' is the axial velocity component function. Figure 3 b is the tangential velocity component function. Figure 3 c is the radial velocity component function;

[0040] Figure 4 The resulting graph shows the inlet condition function under the OP=0.45 operating condition. Figure 4 'a' is the axial velocity component function. Figure 4 b is the tangential velocity component function. Figure 4 c is the radial velocity component function;

[0041] Figure 5 The graph shows the statistical results of the error of the ingress condition function.

[0042] Figure 6 This is a visualization of the prior prediction flow field for a single-channel multi-condition operation under the predicted operating conditions. Figure 6 a represents the pressure prediction field for a single flow channel. Figure 6 b represents the single-channel pressure CFD reference field. Figure 6 c represents the velocity prediction field for a single flow channel. Figure 6 d represents the single-channel velocity CFD reference field;

[0043] Figure 7 A graph showing the statistical results of prior prediction errors for multiple operating conditions in a single flow channel under the conditions to be predicted.

[0044] Figure 8 The image shows a comparison of the residual velocity field of the guide wheel's entire flow path before and after correction under the OP=0.35 condition. Figure 8 'a' represents the direct mapping of the prior velocity field. Figure 8 b is the residual corrected velocity field. Figure 8 c represents the CFD reference velocity field;

[0045] Figure 9 The image shows a comparison of the residual pressure field of the guide wheel's entire flow channel before and after correction under the OP=0.35 condition. Figure 9 a represents the direct mapping of the prior pressure field. Figure 9 b represents the residual correction pressure field. Figure 9 c represents the CFD reference pressure field;

[0046] Figure 10 The error statistics are shown in the figure for directly mapping the a priori flow field and the residual corrected flow field. Detailed Implementation

[0047] 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 and not intended to limit the invention.

[0048] The specific implementation of the present invention will be described in detail below with reference to specific embodiments. It should be noted that, in this embodiment, the target blade cascade is a 236 hydraulic torque converter guide vane cascade, which has 28 periodic channels. The full flow channel of the target blade cascade refers to the fluid region formed by all the flow channels corresponding to the target blade cascade, and is not limited to the entire flow channel of the hydraulic torque converter pump impeller, turbine, and guide vane integrally coupled. The flow field of the full flow channel of the target blade cascade includes the pressure field and velocity field within the fluid region.

[0049] like Figure 1 As shown, an embodiment of the present invention provides a priori guided residual prediction method for the flow field of a hydraulic torque converter cascade, comprising the following steps:

[0050] Step 1: Acquire flow field data for the entire flow channel under known operating conditions of the target blade cascade;

[0051] First, the full-channel flow field data of the target blade cascade under multiple known operating conditions is obtained. The full-channel flow field data is obtained through computational fluid dynamics simulation, acquiring the spatial coordinates, pressure field, velocity field, and inlet cross-section velocity distribution within the entire flow channel of the target blade cascade.

[0052] The target blade cascade is a blade cascade component with a circumferentially periodic flow channel structure in a hydraulic torque converter. The target blade cascade full flow channel refers to the entire flow channel region corresponding to the target blade cascade. The target blade cascade includes a guide wheel blade cascade, a pump wheel blade cascade, or a turbine blade cascade. Specifically, such as... Figure 2 a- Figure 2 As shown in Figure d, the target blade cascade is the guide vane cascade of a 236 hydraulic torque converter. The guide vane cascade has multiple flow channel elements that are repeatedly distributed along the circumference. The single flow channel computational domain is a local flow channel computational domain selected according to the circumferential periodicity of the guide vane cascade. The inlet section of the target blade cascade is located at the inlet position of the entire flow channel of the guide vane cascade and is used to extract the inlet velocity distribution under different operating conditions.

[0053] The known full-channel flow field data is mainly used to extract the inlet boundary information of the target cascade and to provide the full-channel reference flow field of the target cascade during the training stage of the residual correction model. The training data used by the single-channel multi-condition prior prediction model is not obtained by directly extracting the single-channel physical field from the full-channel flow field, but is obtained by conducting single-channel simulation based on the inlet condition function.

[0054] In this embodiment, the known operating conditions are selected as OP=0.30, OP=0.40 and OP=0.50; the operating conditions to be predicted include OP=0.35 and OP=0.45.

[0055] Step 2: Construct the entry condition function;

[0056] Based on the known velocity distribution at the inlet cross-section of the target cascade's entire flow path under known operating conditions, the inlet velocity is decomposed into components, subjected to radial statistics, interpolation, or fitting to obtain an inlet condition function characterizing the changes in the inlet boundary under different operating conditions. This inlet condition function describes the variation of the magnitude, direction, and spatial distribution of the target cascade's inlet velocity with different operating conditions.

[0057] Specifically, inlet velocity sampling points and their spatial coordinates are extracted from the inlet section of the target cascade's entire flow channel. The inlet velocity sampling points include velocity vectors at different radial and circumferential positions on the inlet section.

[0058] Using the rotation axis of the hydraulic torque converter as the axial direction, a cylindrical coordinate description of the inlet section is established. For any velocity sampling point on the inlet section, its radial and circumferential positions are determined based on its spatial coordinates, and the velocity vector at that point is decomposed into axial velocity components, tangential velocity components, and radial velocity components. The axial velocity component characterizes the velocity along the hydraulic torque converter axis, the tangential velocity component characterizes the velocity along the circumferential direction of the inlet section, and the radial velocity component characterizes the velocity along the radial direction of the inlet section.

[0059] To transform the velocity distribution of the entire inlet section of the target cascade into an inlet boundary condition applicable to a single-channel computational domain, radial binning statistics are performed on the inlet section velocity data. Radial binning statistics involve dividing the inlet section into several radial intervals along its radial direction, grouping velocity sampling points falling within the same radial interval into the same bin, and averaging or weighted averaging the velocity components at different circumferential positions within each bin to obtain representative values ​​of the axial, tangential, and radial velocities corresponding to that radial interval. Through this processing, the two-dimensional velocity distribution on the inlet section of the entire inlet section of the target cascade can be transformed into a three-component inlet velocity profile that varies with radial position.

[0060] In this embodiment, let the first The radial representative position corresponding to each radial compartment is: The number of velocity sampling points falling into this sub-bin is Then, the representative values ​​of the axial velocity, tangential velocity, and radial velocity within the radial compartment can be expressed as follows:

[0061]

[0062]

[0063]

[0064] in, , and They represent the first The first radial compartment Axial velocity components, tangential velocity components, and radial velocity components at each sampling point; , and They represent the first Representative values ​​of axial velocity, tangential velocity, and radial velocity corresponding to each radial compartment.

[0065] Furthermore, the three-component inlet velocity profiles obtained from multiple radial bins are interpolated or fitted to obtain inlet condition functions characterizing the changes in the inlet boundary under different operating conditions. For operating condition OP, its inlet condition function can be expressed as:

[0066]

[0067] in, Indicates the radial position of the inlet section. , and These represent the axial velocity component function, tangential velocity component function, and radial velocity component function under the operating condition OP, respectively. This represents the inlet condition function that varies with radial position under operating condition OP.

[0068] In single-channel simulation, based on the radial position of the inlet boundary node of the single-channel, the axial velocity, tangential velocity, and radial velocity at that node are obtained from the inlet condition function of the corresponding working condition. Let the first node be the first node at the inlet boundary of the single-channel. The radial position of each node is Then the inlet velocity component at that node can be expressed as:

[0069]

[0070] in, This represents the axial velocity component at the i-th node of the flow channel inlet boundary under operating condition OP. This represents the tangential velocity component at the i-th node of the flow channel inlet boundary under operating condition OP. This represents the radial velocity component at the i-th node of the flow channel inlet boundary under operating condition OP. This represents the axial velocity component function of the operating condition OP at the radial position r. i The axial velocity calculated at that point, This represents the tangential velocity component function of the operating condition OP at the radial position r. i The tangential velocity calculated at that point, This represents the radial velocity component function of the operating condition OP at the radial position r. i The radial velocity is calculated at the inlet node. Subsequently, based on the circumferential position of the inlet node in the single-channel coordinate system, the aforementioned axial, tangential, and radial velocity components are converted into velocity vector components required by the solver and applied to the single-channel inlet boundary. Thus, the velocity distribution of the entire inlet section of the target cascade is transformed into an inlet velocity boundary condition usable in the single-channel computational domain.

[0071] In this embodiment, OP=0.30, OP=0.40, and OP=0.50 are selected as known working conditions, and corresponding ingress condition functions are constructed for each. For the intermediate working condition OP=0.35 that needs to be predicted, the axial velocity, tangential velocity, and radial velocity at each radial position are interpolated based on the ingress condition functions corresponding to the two known working conditions OP=0.30 and OP=0.40 to generate the ingress condition function corresponding to the OP=0.35 working condition; for OP=0.45, the same method is used to generate the ingress condition function corresponding to the two known working conditions OP=0.40 and OP=0.50.

[0072] Specifically, for any operating condition OP to be predicted, if it lies within two adjacent known operating conditions... and Between, among Then first obtain and The corresponding entry condition function:

[0073]

[0074]

[0075] in, Indicates known working conditions The inlet condition function varies with radial position. Indicates known working conditions The inlet condition function varies with radial position. , and These represent known operating conditions. The axial velocity component function, tangential velocity component function, and radial velocity component function are given below. , and These represent known operating conditions. The axial velocity component function, tangential velocity component function, and radial velocity component function are given.

[0076] Then for each radial position Interpolation calculations are performed on the axial, tangential, and radial velocity components along the working direction, respectively. Taking linear interpolation as an example, the interpolation weights are... Defined as:

[0077]

[0078] The axial velocity component function, tangential velocity component function, and radial velocity component function under the predicted working condition OP are expressed as follows:

[0079]

[0080]

[0081]

[0082] This yields the input condition function corresponding to the working condition to be predicted:

[0083]

[0084] Taking OP=0.35 as an example, it lies between OP=0.30 and OP=0.40, at which point the interpolation weights... =0.5, therefore, the three velocity components at each radial position under the OP=0.35 condition are all obtained by linearly combining the velocity components corresponding to the OP=0.30 and OP=0.40 conditions with equal weights. Similarly, the inlet condition function corresponding to the OP=0.45 condition can be generated using the inlet condition functions of the OP=0.40 and OP=0.50 conditions.

[0085] In other implementations, spline interpolation, polynomial fitting, piecewise cubic Hermite interpolation, or other continuous function fitting methods can be used to construct the inlet condition function for the predicted operating condition. The inlet condition function generated by interpolation or fitting can maintain the radial velocity distribution characteristics and reflect the continuous change law of the inlet flow state between different operating conditions. Subsequently, the generated inlet condition function is applied to the inlet boundary of the corresponding single-channel computational domain for subsequent single-channel simulation calculations and data construction of single-channel multi-operating-condition prior prediction models.

[0086] In this embodiment, MAE represents the mean absolute error, RMSE represents the root mean square error, and R... 2 The coefficient of determination is denoted by nRMSE, which represents the normalized root mean square error. It is used to evaluate the deviation between the inlet condition function, the single-channel prior prediction results, and the residual correction results and the reference data.

[0087] like Figures 3 to 5 As shown, based on the velocity distribution extracted from the inlet section of the target cascade under known operating conditions, component decomposition, radial binning statistics, circumferential averaging, and interpolation or fitting processing are performed on the inlet velocity to obtain the inlet condition function under different operating conditions. Among these, Figure 3 The results of constructing the inlet velocity component function under the OP=0.35 condition are shown. Figure 3 a, Figure 3 b and Figure 3 c shows the axial velocity component function, tangential velocity component function, and radial velocity component function, respectively; Figure 4 The results of constructing the inlet velocity component function under the OP=0.45 condition are shown. Figure 4 a, Figure 4 b and Figure 4 c shows the axial velocity component function, tangential velocity component function, and radial velocity component function, respectively; Figure 5 The error statistics for the inlet velocity component function are shown. (From...) Figures 3 to 5 It can be seen that the inlet condition function can characterize the changes in inlet velocity distribution under different operating conditions and provide inlet boundary conditions for subsequent single-channel simulation and single-channel multi-condition prior prediction model.

[0088] Step 3: Generation of single-channel simulation data and training of single-channel multi-condition prior prediction model;

[0089] A single-channel computational domain is constructed based on the circumferential periodicity of the target cascade. The inlet condition functions corresponding to different operating conditions are applied to the single-channel computational domain as inlet boundary conditions to perform single-channel flow field simulation and obtain single-channel simulation data under multiple operating conditions.

[0090] The single-channel simulation data includes spatial coordinates, pressure field, and velocity field within the single-channel computational domain. This data is used to train a single-channel multi-condition prior prediction model. In this embodiment, the single-channel multi-condition prior prediction model is not trained from the single-channel physical field directly extracted from the full-channel flow field of the target cascade, but rather from single-channel simulation data driven by the inlet condition function.

[0091] Furthermore, a multi-condition prior prediction model for a single-channel flow path with inlet function conditionation is trained based on single-channel simulation data under different operating conditions. This multi-condition prior prediction model takes the single-channel spatial coordinates and inlet condition function as input, and the single-channel pressure field and velocity field as output, to predict the flow field distribution within the single channel of the target cascade under the corresponding operating conditions.

[0092] In this embodiment, the single-channel multi-condition prior prediction model is implemented using a neural network surrogate model with inlet function conditionalization. The model input includes a fixed reference single-channel node table, inlet conditional function encoding, and boundary labels. The fixed reference single-channel node table includes the spatial coordinates and radial position features of the single channel; the boundary labels characterize the inlet, outlet, wall, periodic boundary, or internal fluid region to which the node belongs; and the inlet conditional function encoding characterizes the velocity distribution at the inlet cross-section under the corresponding condition. The model output consists of the pressure components and three-dimensional velocity components at the corresponding node.

[0093] Specifically, the single-channel multi-condition prior prediction model in this embodiment includes a spatial encoding module, an inlet condition encoding module, a feature fusion module, and a flow field decoding module. The spatial encoding module takes the single-channel spatial coordinates, radial position features, and boundary labels as input, and uses multi-scale position feature encoding and a two-layer fully connected network to extract spatial latent features. The inlet condition encoding module takes the operating parameters, inlet reference velocity, and inlet condition function encoding vector as input, and uses low-frequency condition feature encoding and a two-layer fully connected network to extract inlet condition latent features. The feature fusion module concatenates the spatial latent features and the inlet condition latent features to obtain fused features. The flow field decoding module outputs the pressure components and three-dimensional velocity components at the nodes based on the fused features.

[0094] In this embodiment, the fully connected network of the spatial coding module has 128 neurons per layer, outputting 128-dimensional spatial latent features; the fully connected network of the inlet conditional coding module has 128 neurons per layer, outputting 64-dimensional inlet conditional latent features; the flow field decoding module uses a residual decoding network with a width of 256, containing four residual connection modules. Each residual connection module consists of a fully connected layer, a normalization layer, and a nonlinear activation function. Finally, the pressure component and three-dimensional velocity component are output through a linear output layer. The nonlinear activation function is the SiLU function, and the normalization layer is LayerNorm.

[0095] During model training, the pressure and velocity fields from single-channel simulation data are used as supervisory data to construct a supervisory loss function composed of pressure and velocity error terms, and the Adam optimizer is used for parameter updates. Before training, the single-channel spatial coordinates, inlet condition function encoding, and output flow field variables are dimensionless or normalized. It should be noted that the above network structure, number of layers, number of neurons, feature encoding method, activation function, normalization method, and training parameters are only one possible implementation in this embodiment. In other embodiments, adjustments can be made according to the data scale and prediction requirements, and this invention does not limit these adjustments.

[0096] After training, the single-channel multi-condition prior prediction model is used to generate single-channel predicted flow fields under known or unpredicted conditions. In the subsequent residual correction model training phase, the parameters of the single-channel multi-condition prior prediction model remain unchanged and are used as a fixed model to construct the directly mapped prior flow field.

[0097] Under the predicted operating condition OP=0.35, the single-channel multi-condition prior prediction model can obtain the predicted pressure and velocity fields of the single channel based on the inlet condition function. Among these, Figure 6 a to Figure 6 d represents the visualization result of the predicted flow field in a single channel under the OP=0.35 condition. Figure 7 This presents the statistical results of single-channel prediction errors under the predicted operating conditions. Combined with... Figure 6and Figure 7 This demonstrates that the single-channel multi-condition prior prediction model has the ability to predict intermediate conditions that were not trained, and can provide single-channel prior information for subsequent direct mapping of the prior flow field construction and residual correction.

[0098] Step 4: Full-circuit rotation mapping of single-channel prediction results;

[0099] Using the trained single-channel multi-condition prior prediction model, the single-channel predicted flow field under known or unpredicted conditions is obtained. Then, based on the circumferential position and rotation angle of the target blade cascade, the single-channel predicted flow field is periodically mapped to the entire flow channel of the target blade cascade to obtain the directly mapped prior flow field.

[0100] In this embodiment, the direct mapping of the prior flow field can be constructed using a reverse mapping method oriented towards the target cascade full-channel node. Specifically, firstly, the circumferential position and rotation angle of the target cascade full-channel node are identified, and the full-channel node is reverse-mapped to the reference single-channel coordinate system according to the corresponding period angle. Then, the reverse-mapped single-channel coordinate features and the inlet condition function of the corresponding operating condition are input into the trained single-channel multi-operating-condition prior prediction model to obtain the pressure prior value and velocity prior value of the node in the reference single-channel coordinate system. Finally, the pressure prior value and velocity prior value are mapped back to the target cascade full-channel coordinate system, where the pressure remains unchanged as a scalar, and the velocity is transformed by directional rotation according to the corresponding period angle.

[0101] If the target cascade has If there are 1 periodic channel, then the circumferential angle between adjacent periodic channels is:

[0102]

[0103] For the Each periodic channel has a rotation angle of:

[0104]

[0105] In this embodiment, the 236 hydraulic torque converter guide vane cascade has 28 periodic channels, i.e. =28.

[0106] In the target blade cascade full flow channel The node coordinates within each periodic channel are: Its coordinates after being mapped back to the reference single-channel coordinate system are: Then it can be expressed as:

[0107]

[0108] in, The angle of rotation about the axis of the hydraulic torque converter is represented by -θ. kThe rotation matrix is ​​used to reverse map the node coordinates in the k-th period channel to the reference single-channel coordinate system; This is the inverse transformation, used to subsequently map the velocity vector in the reference single-channel coordinate system back to the k-th cycle channel. In this embodiment, if the hydraulic torque converter axis is taken as the x-axis, the rotation matrix can be expressed as:

[0109]

[0110] The coordinates after reverse mapping and the corresponding input condition function Input the single-channel multi-condition prior prediction model to obtain the pressure prior value in the reference single-channel coordinate system. and velocity prior value After mapping back to the k-th period channel, the pressure and velocity can be expressed as follows:

[0111]

[0112]

[0113] in, and These represent the pressure prior value and velocity prior value of the corresponding node in the k-th cycle channel, respectively.

[0114] In this way, the local flow field information output by the single-channel multi-condition prior prediction model can be extended to the entire flow channel of the target cascade, forming a directly mapped prior flow field. This directly mapped prior flow field retains the single-channel flow field prediction results obtained based on the inlet condition function, and maintains the consistency of spatial coordinates and velocity directions between different periodic channels through pressure scalar mapping and velocity vector rotation transformation, which can be used as the prior input for the subsequent residual correction model.

[0115] Step 5: Correction of full-channel residuals based on prior guidance;

[0116] Due to inlet condition function fitting errors, single-channel prediction errors, and coordinate and vector transformation errors during the periodic mapping process, residual deviations may still exist between the directly mapped prior flow field and the reference flow field of the target cascade's entire flow channel under the corresponding operating conditions. To compensate for this deviation, this embodiment constructs a residual correction model, using the directly mapped prior flow field as the prior input, and learns its residuals relative to the reference flow field of the target cascade's entire flow channel.

[0117] Let the directly mapped prior flow field be... The reference flow field of the target cascade's entire flow channel obtained by CFD simulation under the corresponding operating condition is: Then the residual supervision during the training phase can be expressed as:

[0118]

[0119] in, The residual field represents the flow field between the reference flow field of the entire flow channel of the target cascade and the directly mapped prior flow field, and is used as supervision data during the training phase of the residual correction model. Including pressure residuals and velocity residuals, it can be expressed as:

[0120]

[0121] in, This represents the pressure residual between the reference pressure field of the entire flow channel of the target blade cascade and the directly mapped a priori pressure field. This represents the velocity residual between the reference velocity field of the entire flow channel of the target blade cascade and the directly mapped prior velocity field.

[0122] In this embodiment, the inputs to the residual correction model include the directly mapped prior flow field, the spatial coordinates of the entire flow channel of the target blade cascade, and the inlet condition function; the outputs include pressure residuals and velocity residuals. Let the spatial coordinates of the entire flow channel of the target blade cascade be... The entry condition function is The residual correction model is The output of the residual correction model can then be expressed as:

[0123]

[0124] in, These are the pressure and velocity residuals predicted by the residual correction model. When training the residual correction model, the parameters of the single-channel multi-condition prior prediction model are kept constant, allowing the residual correction model to primarily learn the systematic deviation between the directly mapped prior flow field and the reference flow field of the target cascade's entire flow channel.

[0125] In the prediction phase, the directly mapped prior flow field corresponding to the operating condition to be predicted is first obtained according to the aforementioned periodic mapping method. Then, this directly mapped prior flow field, the spatial coordinates of the entire flow channel of the target cascade, and the inlet condition function are input into the trained residual correction model to obtain the prediction residual. The corrected prediction result of the flow field of the entire flow channel of the target cascade can be expressed as:

[0126]

[0127] in, This represents the predicted flow field across the entire target cascade after residual correction, including the corrected pressure and velocity fields. Further, if the pressure and velocity fields are expressed separately, the corrected pressure and velocity fields can be expressed as:

[0128]

[0129]

[0130] in, and These represent the pressure and velocity in the direct mapping of the prior flow field, respectively. and These represent the pressure residual and velocity residual output by the residual correction model, respectively. and These represent the pressure field and velocity field after residual correction, respectively.

[0131] The target cascade full-channel reference flow field is only used to construct residual supervision data during the residual correction model training phase; during the prediction phase of the operating condition to be predicted, it is not necessary to obtain the target cascade full-channel reference flow field corresponding to that operating condition. In this embodiment, for the operating condition to be predicted, its target cascade full-channel reference flow field is only used for error evaluation and effect verification after the prediction is completed, and does not participate in the model input and residual correction process for that operating condition.

[0132] like Figure 8 As shown, the predicted velocity field of a single flow channel is periodically mapped according to the period angle of the target blade cascade to obtain the directly mapped prior velocity field; further, the directly mapped prior velocity field is corrected by a residual correction model to obtain the corrected predicted velocity field of the entire flow channel of the guide wheel. Figure 8 a- Figure 8 As can be seen, the velocity field after residual correction is closer to the CFD velocity field than the directly mapped prior velocity field.

[0133] like Figure 9 As shown, the predicted pressure field of a single flow channel is mapped to the entire flow channel of the target blade cascade according to the period angle of the target blade cascade, resulting in a directly mapped prior pressure field. Further correction of the directly mapped prior pressure field is achieved using a residual correction model, yielding the corrected predicted pressure field for the entire flow channel of the guide vane. Figure 9 a- Figure 9 As can be seen from c, the pressure field after residual correction is closer to the CFD pressure field in terms of overall distribution, indicating that the residual correction model can compensate for the prediction bias in the direct mapping of the prior pressure field.

[0134] like Figure 10 As shown, error statistics were performed on the directly mapped prior flow field and the residual corrected flow field under known and predicted operating conditions. The statistical results show that the errors of the pressure field and velocity field after residual correction are lower than those of the directly mapped prior flow field, indicating that the method of the present invention can effectively reduce the residual deviation of the single-channel predicted flow field after periodic mapping, and maintain a good prediction effect of the target cascade full-channel flow field under known and predicted operating conditions.

[0135] 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 of priori guided residual prediction of a fluid torque converter cascade flow field, characterized by, Includes the following steps: Step 1: Obtain the full-channel flow field data of the target blade cascade under multiple known operating conditions; the full-channel flow field data includes the spatial coordinates, pressure field, velocity field, and inlet cross-section velocity distribution within the full-channel of the target blade cascade. Step 2: Based on the inlet cross-sectional velocity distribution of the target blade cascade under known operating conditions, perform component decomposition, radial statistics, interpolation or fitting on the inlet velocity to obtain the inlet condition function used to characterize the changes in the inlet boundary under different operating conditions; Step 3: Construct a single-channel computational domain based on the circumferential periodicity of the target cascade, and apply the inlet condition functions corresponding to different operating conditions as inlet boundary conditions to the single-channel computational domain to perform single-channel flow field simulation and obtain single-channel simulation data under multiple operating conditions; train a single-channel multi-operating-condition prior prediction model with inlet function conditionalization based on the single-channel simulation data. Step 4: Use the trained single-channel multi-condition prior prediction model to obtain the single-channel predicted flow field under known or predicted conditions. Based on the circumferential position and rotation angle of the target blade cascade, periodically map the single-channel predicted flow field to the entire flow channel of the target blade cascade to obtain the directly mapped prior flow field. Step 5: Construct a residual correction model, using the direct mapping prior flow field, the spatial coordinates of the target cascade's entire flow channel, and the inlet condition function as inputs, and the pressure residual and velocity residual as outputs, to learn the residual of the direct mapping prior flow field relative to the reference flow field of the target cascade's entire flow channel; In the prediction phase, the direct-mapped prior flow field corresponding to the working condition to be predicted is input into the trained residual correction model to obtain the prediction residual. The prediction residual is then superimposed with the direct-mapped prior flow field to obtain the corrected prediction result of the full flow field of the target cascade.

2. The prior-guided residual prediction method for the flow field of a hydraulic torque converter blade cascade according to claim 1, characterized in that, The target blade cascade is a blade cascade component with a circumferential periodic flow channel structure in a hydraulic torque converter, including a guide wheel blade cascade, a pump wheel blade cascade, or a turbine blade cascade. The target blade cascade full flow channel refers to all flow channel regions corresponding to the target blade cascade, and the single flow channel calculation domain is a local flow channel calculation domain selected from the target blade cascade full flow channel according to the circumferential periodicity of the target blade cascade.

3. The a priori guided residual prediction method of a fluid torque converter cascade flow field of claim 1, wherein, In step 2, the method for constructing the entry condition function includes: The inlet velocity sampling points and their spatial coordinates are extracted from the inlet section of the entire flow channel of the target cascade. The inlet velocity sampling points include velocity vectors at different radial and circumferential positions on the inlet section. Using the rotation axis of the hydraulic torque converter as the axial direction, a cylindrical coordinate description of the inlet section is established. For any velocity sampling point on the inlet section, the radial and circumferential positions are determined according to its spatial coordinates, and the velocity vector of that point is decomposed into axial velocity components, tangential velocity components, and radial velocity components. The inlet section is divided into several radial intervals along the radial direction of the inlet section for radial binning statistics. Velocity sampling points whose radial positions fall within the same radial interval are grouped into the same bin. The velocity components at different circumferential positions in each bin are averaged or weighted averaged to obtain the representative values ​​of axial velocity, tangential velocity and radial velocity corresponding to the radial interval. Thus, the two-dimensional velocity distribution on the inlet section is transformed into a three-component inlet velocity profile that varies with the radial position. Interpolation or fitting is performed on the three-component inlet velocity profile obtained from multiple radial bins to obtain the inlet condition function used to characterize the changes in the inlet boundary under different working conditions.

4. The a priori guided residual prediction method of a fluid torque converter cascade flow field according to claim 3, wherein, In step 2, for the working condition to be predicted located between two adjacent known working conditions, the ingress condition function corresponding to the working condition to be predicted is generated in the following way: First, obtain the inlet condition functions corresponding to the two adjacent known working conditions; then, for each radial position, perform interpolation calculations on the axial velocity component, tangential velocity component and radial velocity component in the working direction to obtain the velocity component functions under the working condition to be predicted, thereby forming the inlet condition function corresponding to the working condition to be predicted.

5. The method of priori guided residual prediction of a fluid torque converter cascade flow field of claim 4, wherein, The single-channel multi-condition prior prediction model includes a spatial coding module, an inlet condition coding module, a feature fusion module, and a flow field decoding module. The spatial coding module takes single-channel spatial coordinates, radial position features and boundary labels as input, and uses multi-scale position feature coding and a two-layer fully connected network to extract spatial latent features. The entry condition coding module takes the operating parameters, entry reference speed and entry condition function coding vector as input, and uses low-frequency condition feature coding and a two-layer fully connected network to extract the latent features of the entry condition. The feature fusion module is used to concatenate spatial latent features and entry condition latent features to obtain fused features; The flow field decoding module is used to output the pressure component and three-dimensional velocity component at the node based on the fusion features.

6. The a priori guided residual prediction method of a fluid torque converter cascade flow field of claim 5, wherein, In step 4, the periodic mapping adopts a reverse mapping method: Identify the circumferential position and rotation angle of the target blade cascade's full-channel node, and reverse map the full-channel node to the reference single-channel coordinate system according to the corresponding period angle. Input the reverse-mapped single-channel coordinate features and the inlet condition function of the corresponding working condition into the trained single-channel multi-working-condition prior prediction model to obtain the pressure prior value and velocity prior value of the node in the reference single-channel coordinate system. Map the pressure prior value and velocity prior value back to the target blade cascade's full-channel coordinate system, where the pressure remains unchanged as a scalar and the velocity is transformed by directional rotation according to the corresponding period angle.

7. The a priori guided residual prediction method of a fluid torque converter cascade flow field of claim 6, wherein, In step 5, the input of the residual correction model also includes the pressure field and velocity field in the direct mapping prior flow field; when training the residual correction model, the parameters of the single-channel multi-condition prior prediction model are kept unchanged, so that the residual correction model learns the systematic deviation between the direct mapping prior flow field and the reference flow field of the target blade cascade in the entire flow channel.

Images