A centrifugal pump blade optimization method and system fusing two-dimensional feature extraction and machine learning

Abstract

The application belongs to the technical field of centrifugal pump optimization, and discloses a centrifugal pump blade optimization method and system fusing two-dimensional feature extraction and machine learning, which comprises the following steps: obtaining profiles by slicing blades at different blade heights, converting three-dimensional calculation of the blades to two-dimensional calculation on the slices, realizing substantial shortening of time required for sample numerical calculation, and providing the possibility for large-scale sample calculation; and respectively optimizing and designing each slice profile based on a machine learning method, finally fitting and restoring each optimized profile to three dimensions, and realizing performance improvement of the centrifugal pump. The application can improve the precision of the machine learning model through large-scale sample calculation, and fast numerical calculation is convenient for realizing optimization design of more targets and more design variables.

Description

Technical Field

[0001] This invention belongs to the field of centrifugal pump optimization technology, specifically relating to a centrifugal pump blade optimization method and system that integrates two-dimensional feature extraction and machine learning. Background Technology

[0002] Centrifugal pumps, as a widely used general-purpose machine, are common in all aspects of production and daily life, and play an increasingly important role in high-end applications such as aerospace and nuclear industry. In many fields such as water transfer and energy storage, centrifugal pumps are increasingly characterized by high power and high head, and the performance requirements are becoming increasingly stringent. However, the design of centrifugal pumps relies heavily on experience. Therefore, to meet these higher requirements, optimized design has become a crucial and indispensable part of the hydraulic design of centrifugal pumps.

[0003] Currently, optimization design is primarily based on computational fluid dynamics (CFD) techniques. As performance requirements increase, the optimization objectives also grow, leading to a significant increase in the number of design variables and consequently, a substantial increase in the required sample size. Although advancements in computer technology have shortened the computation time for individual samples, this remains a huge time cost when dealing with dozens, hundreds, or even thousands of samples, severely reducing the efficiency of optimization design. This is true even for steady-state calculations; the time cost for transient calculations increases exponentially. Therefore, it is necessary to find a method that can significantly improve the efficiency of numerical computation for individual samples. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a centrifugal pump blade optimization method and system that integrates two-dimensional feature extraction and machine learning. By slicing the blade at different blade heights to reduce dimensionality, and by accurately fitting the functional relationship between variables and the target based on machine learning methods, the centrifugal pump can be optimized efficiently.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A centrifugal pump blade optimization method integrating two-dimensional feature extraction and machine learning, the method comprising:

[0007] The blade profile was obtained by slicing the centrifugal pump blade at different blade heights.

[0008] The contours of each slice are optimized using machine learning methods, and finally the optimized contours are fitted and restored into three-dimensional blades.

[0009] Numerical calculations are performed on the reconstructed 3D blades. If the performance meets the preset conditions, the results are output; otherwise, the design is re-optimized until optimization is completed.

[0010] Preferably, obtaining the blade profile by slicing the centrifugal pump blade at different blade heights includes:

[0011] Calculate the initial flow field, analyze the flow characteristics, and determine the optimization objective;

[0012] Based on the optimization objective, the blade profile is obtained by intersecting the flow surface with the blade at different blade heights, and the design variables are determined.

[0013] Preferably, the optimization design of each slice contour based on machine learning methods, and the final fitting and restoration of each optimized contour into a three-dimensional blade, includes:

[0014] The dynamic modal decomposition analysis method is used to extract the velocity and pressure information at the selected blade height position in the initial flow field to obtain the boundary conditions for two-dimensional calculation;

[0015] Determine the range of variation of the design variables, and use experimental design methods to obtain a number of sample points that meet the preset requirements within the design space;

[0016] Utilize automated numerical calculation methods to complete numerical calculations for sample points and improve the sample database;

[0017] The selected machine learning model is trained using the improved sample library;

[0018] The optimal combination of design variables is obtained by using intelligent optimization algorithms to solve the trained machine learning model.

[0019] The obtained optimal contour combination is restored into a three-dimensional blade.

[0020] Preferably, the blade outline is restored to a three-dimensional blade by successively fitting the outlines at different blade heights to form a cylindrical surface, and then the two ends are closed.

[0021] The present invention also provides a centrifugal pump blade optimization system that integrates two-dimensional feature extraction and machine learning. The system is used to implement the method described in any one of them, and the system includes: an acquisition module, an optimization module, and an iteration module.

[0022] The acquisition module is used to obtain the blade profile by slicing the centrifugal pump blade at different blade heights.

[0023] The optimization module is used to optimize the design of each slice contour based on machine learning methods, and finally fit and restore each optimized contour into a three-dimensional blade.

[0024] The iterative module is used to perform numerical calculations on the restored three-dimensional blade. If the performance meets the preset conditions, the result is output; otherwise, the design is re-optimized until the optimization is completed.

[0025] Preferably, the acquisition module includes: an optimization target determination unit and a blade profile determination unit;

[0026] The optimization target determination unit is used to calculate the initial flow field, analyze the flow characteristics, and determine the optimization target;

[0027] The blade profile determination unit is used to obtain the blade profile by intersecting the flow surface with the blade at different blade heights based on the optimization objective, and to determine the design variables.

[0028] Preferably, the optimization module includes: a boundary condition calculation unit, a sample point acquisition unit, a sample library improvement unit, a model training unit, a solution unit, and a restoration unit;

[0029] The boundary condition calculation unit is used to extract the velocity and pressure information at the selected blade height position in the initial flow field using the dynamic modal decomposition analysis method, and obtain the boundary conditions for two-dimensional calculation.

[0030] The sample point acquisition unit is used to determine the transformation range of the design variables and to acquire a number of sample points that meet the preset requirements within the design space using experimental design methods.

[0031] The sample library improvement unit is used to complete the numerical calculation of sample points and improve the sample library using an automatic numerical calculation method.

[0032] The model training unit is used to train the selected machine learning model using the improved sample library.

[0033] The solving unit is used to solve the trained machine learning model using intelligent optimization algorithms to obtain the optimal combination of design variables;

[0034] The restoration unit is used to restore the obtained optimal contour combination into a three-dimensional blade.

[0035] Preferably, the blade outline is restored to a three-dimensional blade by successively fitting the outlines at different blade heights to form a cylindrical surface, and then the two ends are closed.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] Existing optimization methods primarily rely on three-dimensional computation for sample numerical calculations, which are time-consuming. This invention reduces three-dimensional numerical computation to two dimensions by slicing the three-dimensional leaf, significantly shortening the computation time and enabling computation on ultra-large-scale sample data. This allows for more precise training of machine learning models and more accurate performance predictions. Furthermore, the rapid numerical computation facilitates the optimization design of more objectives and more design variables. Attached Figure Description

[0038] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart illustrating the process of the method described in this embodiment of the invention;

[0040] Figure 2 This is a schematic diagram of a blade slice in an embodiment of the present invention;

[0041] Figure 3 This is a comparison chart of the optimization results of different slices in the embodiments of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] Example 1

[0045] like Figure 1 As shown in the figure, this invention discloses a centrifugal pump blade optimization method that integrates two-dimensional feature extraction and machine learning. The method includes:

[0046] The blade profile was obtained by slicing the centrifugal pump blade at different blade heights.

[0047] The contours of each slice are optimized using machine learning methods, and finally the optimized contours are fitted and restored into three-dimensional blades.

[0048] Numerical calculations are performed on the reconstructed 3D blades. If the performance meets the preset conditions, the results are output; otherwise, the design is re-optimized until optimization is completed.

[0049] The specific implementation process is as follows: (1) Initial calculation. First, use 3D software such as UG to model the hydraulic system, and use commercial CFD software such as ANSYS to perform mesh generation and numerical simulation to complete the initial flow field calculation. By observing the performance indicators such as head and efficiency monitored in the calculation and the internal flow conditions, analyze the flow characteristics, check the flow loss distribution, and clarify the direction to be improved in combination with actual needs, so as to determine the optimization target. Among them, head H and efficiency η can be calculated according to the following formula:

[0050]

[0051] In the formula, p tot2 p tot1 These are the total pressure at the inlet and outlet, respectively.

[0052]

[0053] In the formula, M is the sum of the torques of the impeller blades, and ω is the angular velocity of the impeller rotation.

[0054] (2) Model Slicing. Flow surfaces are generated along the flow direction at different blade heights (span = 0.25, span = 0.5, span = 0.75, etc.) within the model flow channel. The intersection lines of each flow surface and the blades are extracted to obtain the profile. Bézier curves are used to constrain the profile, and the coordinates of the control points are determined as design variables. The range of variation is then defined. The optimization problem can then be described as:

[0055] min f(x1, x2, x3, ...)

[0056] subject to x l <x i <x u i = 1, 2, 3…

[0057] In the formula, x l x u These are the upper and lower boundaries of the design variable, respectively.

[0058] (3) Boundary condition extraction. Dynamic mode decomposition and other analysis methods are used to read the velocity v, pressure p and other information at the selected blade height position in the initial flow field, and the obtained velocity and pressure information is used as the boundary conditions for two-dimensional calculation.

[0059] (4) Experimental design. Experimental design methods are used to sample within a range to obtain a sufficient number of sample points.

[0060] (5) Two-dimensional numerical calculation. An automatic numerical calculation program is used to perform numerical calculations on the sample points, thus improving the sample library. The automatic numerical calculation program controls the CFD software through scripts to realize the automatic numerical simulation process of multiple samples, thereby improving the sample library more quickly.

[0061] (6) Model Training. The selected machine learning model is trained using the improved sample database. The sample database is divided into training, validation, and test sets at proportions of 75%, 15%, and 15% respectively, to ensure the machine learning model achieves the required prediction accuracy (regression coefficient R0). 2 >0.9).

[0062] (7) Algorithm optimization. The model is solved using intelligent optimization algorithms. The optimal solution is found by continuously searching within the design space, thus obtaining the optimal combination of design variables.

[0063] (8) 3D reconstruction and calculation. The obtained optimal contour combination is reconstructed into a 3D blade by directional sweeping, and the reconstructed 3D model is numerically calculated again.

[0064] (9) Performance comparison. Compare the performance with the initial model. If the ideal conditions are met, output the results; otherwise, return to the experimental design and repeat the process until optimization is completed.

[0065] In this embodiment, the number of flow surfaces and contours mentioned in (2) can be selected as needed according to actual conditions.

[0066] In this embodiment, if the flow surface described in (2) is a curved surface, it can be cut off and the contour can be converted into a plane using methods such as conformal transformation.

[0067] In this embodiment, the number of sample points mentioned in (4) can be determined comprehensively based on the optimization objective and the number of design variables, taking into account the requirements of the model used.

[0068] In this embodiment, the experimental design method described in (4) can be selected from classic experimental design methods such as orthogonal experiment and Latin hypercube experimental design.

[0069] In this embodiment, the automatic numerical calculation program described in (5) is mainly implemented by controlling numerical calculation software scripts, which can be written in programming languages ​​such as Python.

[0070] In this embodiment, the machine learning model mentioned in (6) can be selected from models such as artificial neural networks (ANN) and support vector machines (SVM).

[0071] In this embodiment, the model quality mentioned in (6) can be expressed using the linear regression coefficient R. 2 An evaluation will be conducted, and if the standard is met, the next step will be taken.

[0072] In this embodiment, the intelligent optimization algorithm described in (7) can be selected from algorithms such as genetic algorithm (GA), particle swarm optimization (PSO), and gravity search algorithm (GSA).

[0073] In this embodiment, the process of restoring the blade profile into a three-dimensional model as described in (8) can be achieved by sequentially fitting the profiles at different blade heights to form a cylindrical surface, and then closing both ends.

[0074] In this embodiment, if the optimization result is not ideal as described in (9), we can also return to the model training stage to retrain and then seek optimization again.

[0075] In this embodiment, the slice optimization process at each leaf height in (4) to (7) needs to be carried out simultaneously.

[0076] Example 2

[0077] In this embodiment, the model pump is a vertical centrifugal pump, and the specific process of optimization using the method described in this invention is as follows.

[0078] (1) The initial calculation of the flow field was completed using commercial CFD software. The results showed that the hydraulic efficiency under the design conditions was 83%, which was used as the optimization target.

[0079] (2) Slices were made using the flow surface at a blade height of 6 mm from the front and rear cover plates, namely slices A and B, as shown. Figure 2 As shown, the blade outline is obtained.

[0080] (3) Analyze the pressure information on the slice in the initial flow field to obtain the boundary conditions.

[0081] (4) The shape of the slice is controlled by applying constraints to the contour lines on the slice. Twenty coordinate values ​​of 10 control points are selected as design variables, and their range of variation is determined. Two hundred sample points are obtained in the design space using the Latin hypercube experimental design method.

[0082] (5) Use an automatic numerical calculation program written in Python to complete the numerical calculation of sample points and improve the sample library.

[0083] (6) The artificial neural network model was trained using the improved sample database. After testing, the regression coefficient R... 2 All values ​​are greater than 0.93, which meets the engineering requirements.

[0084] (7) The particle swarm optimization algorithm is used to solve the model and obtain the optimal combination of design variables.

[0085] (8) By successively fitting the contours on slices A and B ( Figure 3 A cylindrical surface is formed, and then the two ends are closed to restore it to a three-dimensional blade. Numerical calculations are then performed on the restored three-dimensional model.

[0086] (9) Compared with the performance of the initial model, the computational efficiency of the optimized model reaches 87%, which meets the requirements.

[0087] In the above processes (4) to (7), contour optimization on slices A and B is performed simultaneously.

[0088] If the performance does not meet the standards in the above process (9), the experimental design can be returned to and the process repeated until optimization is completed.

[0089] Example 3

[0090] The present invention also provides a centrifugal pump blade optimization system that integrates two-dimensional feature extraction and machine learning. The system is used to implement the method described in any one of them, and the system includes: an acquisition module, an optimization module, and an iteration module.

[0091] The acquisition module is used to obtain the blade profile by slicing the centrifugal pump blade at different blade heights.

[0092] The optimization module is used to optimize the design of each slice contour based on machine learning methods, and finally fit and restore each optimized contour into a three-dimensional blade.

[0093] The iteration module is used to perform numerical calculations on the reconstructed 3D blade. If the performance meets the preset conditions, the result is output; otherwise, the design is re-optimized until optimization is completed.

[0094] In this embodiment, the acquisition module includes: an optimization target determination unit and a blade profile determination unit;

[0095] The optimization objective determination unit is used to calculate the initial flow field, analyze flow characteristics, and determine the optimization objective.

[0096] The blade profile determination unit is used to obtain the blade profile at different blade heights by utilizing the intersection of the flow surface and the blade, based on the optimization objective, and to determine the design variables.

[0097] In this embodiment, the optimization module includes: a boundary condition calculation unit, a sample point acquisition unit, a sample library improvement unit, a model training unit, a solution unit, and a restoration unit;

[0098] The boundary condition calculation unit is used to extract the velocity and pressure information at the selected blade height position in the initial flow field using the dynamic modal decomposition analysis method, and obtain the boundary conditions for two-dimensional calculation.

[0099] The sample point acquisition unit is used to determine the range of change of the design variables and to acquire a number of sample points that meet the preset requirements within the design space using experimental design methods.

[0100] The sample library improvement unit is used to complete the numerical calculation of sample points and improve the sample library using automatic numerical calculation methods.

[0101] The model training unit is used to train the selected machine learning model using the improved sample library.

[0102] The solution unit is used to solve the trained machine learning model using intelligent optimization algorithms to obtain the optimal combination of design variables;

[0103] The reconstruction unit is used to reconstruct the obtained optimal contour combination into a three-dimensional blade.

[0104] In this embodiment, the blade outline is restored to a three-dimensional blade by successively fitting the outlines at different blade heights to form a cylindrical surface, and then the two ends are closed.

[0105] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for optimizing centrifugal pump blades by integrating two-dimensional feature extraction and machine learning, characterized in that, The method includes: The blade profile was obtained by slicing the centrifugal pump blade at different blade heights. The contours of each slice are optimized using machine learning methods, and finally the optimized contours are fitted and restored into three-dimensional blades. Numerical calculations are performed on the reconstructed 3D blades. If the performance meets the preset conditions, the results are output; otherwise, the design is re-optimized until optimization is completed. The blade profile is obtained by slicing the centrifugal pump blade at different blade heights, including: Calculate the initial flow field, analyze the flow characteristics, and determine the optimization objective; Based on the optimization objective, the blade profile is obtained by intersecting the flow surface with the blade at different blade heights, and the design variables are determined. Based on machine learning methods, the contours of each slice are optimized and designed separately. Finally, the optimized contours are fitted and restored to form a three-dimensional blade, including: The dynamic modal decomposition analysis method is used to extract the velocity and pressure information at the selected blade height position in the initial flow field to obtain the boundary conditions for two-dimensional calculation; Determine the range of variation of the design variables, and use experimental design methods to obtain a number of sample points that meet the preset requirements within the design space; Utilize automated numerical calculation methods to complete numerical calculations for sample points and improve the sample database; The selected machine learning model is trained using the improved sample library; The optimal combination of design variables is obtained by solving the trained machine learning model using intelligent optimization algorithms. The obtained optimal contour combination is restored into a three-dimensional blade.

2. The method according to claim 1, characterized in that, The blade outline is restored to a three-dimensional blade by successively fitting the outline at different blade heights to form a cylindrical surface, and then closing both ends.

3. A centrifugal pump blade optimization system integrating two-dimensional feature extraction and machine learning, said system being used to implement the method described in any one of claims 1-2, characterized in that, The system includes: an acquisition module, an optimization module, and an iteration module; The acquisition module is used to obtain the blade profile by slicing the centrifugal pump blade at different blade heights. The optimization module is used to optimize the design of each slice contour based on machine learning methods, and finally fit and restore each optimized contour into a three-dimensional blade. The iterative module is used to perform numerical calculations on the restored three-dimensional blade. If the performance meets the preset conditions, the result is output; otherwise, the design is re-optimized until optimization is completed. The acquisition module includes: an optimization target determination unit and a blade profile determination unit; The optimization target determination unit is used to calculate the initial flow field, analyze the flow characteristics, and determine the optimization target; The blade profile determination unit is used to obtain the blade profile by the intersection of the flow surface and the blade at different blade heights based on the optimization objective, and to determine the design variables. The optimization module includes: a boundary condition calculation unit, a sample point acquisition unit, a sample library improvement unit, a model training unit, a solution unit, and a restoration unit; The boundary condition calculation unit is used to extract the velocity and pressure information at the selected blade height position in the initial flow field using the dynamic modal decomposition analysis method, and obtain the boundary conditions for two-dimensional calculation. The sample point acquisition unit is used to determine the transformation range of the design variables and to acquire a number of sample points that meet the preset requirements within the design space using experimental design methods. The sample library improvement unit is used to complete the numerical calculation of sample points and improve the sample library using an automatic numerical calculation method. The model training unit is used to train the selected machine learning model using the improved sample library. The solving unit is used to solve the trained machine learning model using intelligent optimization algorithms to obtain the optimal combination of design variables; The restoration unit is used to restore the obtained optimal contour combination into a three-dimensional blade.

4. The system according to claim 3, characterized in that, The blade outline is restored to a three-dimensional blade by successively fitting the outline at different blade heights to form a cylindrical surface, and then closing both ends.

Images