Centrifugal pump pressure pulsation optimization method and system based on point cloud neural network

Abstract

The invention discloses a centrifugal pump pressure pulsation optimization method and system based on a point cloud neural network, and belongs to the technical field of fluid machinery optimization design. Firstly, a database containing structure parameters and working conditions is established; structure, flow field point cloud and pressure pulsation data are acquired through CFD simulation and standardized; constructing a point cloud neural network fused with a structure parameter attention mechanism, adding attention branches in a PointNet architecture, distributing structure parameter weights through a self-attention mechanism, extracting weighted structure features, splicing the weighted structure features with point cloud geometry-flow coupling features, and inputting MLP to predict pressure pulsation; after training, constructing a parameter sensitivity matrix based on attention weight distribution and prediction variability; selecting a high-sensitivity parameter as an optimization variable, taking minimization of pressure pulsation of a key monitoring point as a target, and adopting a gradient descent method for iteration under the constraint that lift and efficiency are not lower than reference values; and finally, a low-pulse scheme is output through CFD closed-loop verification. The method is short in time consumption, high in efficiency, good in robustness and generalizability and suitable for optimization of different pump bodies.

Description

Technical Field

[0001] This invention belongs to the field of fluid machinery optimization design technology, specifically relating to a centrifugal pump pressure pulsation optimization method and system based on point cloud neural networks. Background Technology

[0002] Centrifugal pumps are common impeller-type fluid transport devices widely used in various engineering fields such as petrochemicals, power energy, shipping, and metallurgy. With the continuous improvement of system efficiency, stability, and reliability, the unsteady flow characteristics inside centrifugal pumps and the resulting pressure pulsation problems have gradually attracted attention. Pressure pulsation refers to the periodic fluctuations in fluid pressure caused by factors such as flow disturbances, periodic excitation, and geometric asymmetry during impeller rotation. This phenomenon not only affects the pump's hydraulic performance but also causes noise, vibration, and fatigue damage, ultimately impacting the stable operation and service life of the equipment.

[0003] Traditional optimization for pressure pulsations primarily relies on CFD (Computational Fluid Dynamics) simulation techniques. This involves scanning the impeller's geometric parameters one by one, comparing the pressure response values ​​under different structural conditions, and using response surface methodology (RSM) to find the relatively optimal design. However, RSM is slow in real-time optimization and can only provide good optimization results within a certain approximate range, failing to adapt to a large design space. Other methods, such as genetic algorithms, particle swarm optimization, and multi-objective optimization, cover all possible parameter combinations, are complex processes, and struggle to directly output optimized solutions end-to-end. Furthermore, traditional methods lack sufficient flexibility and generalization ability when dealing with high-dimensional parameter spaces and strongly coupled nonlinear systems, making it difficult to meet the demands for rapid prediction and iterative optimization in complex engineering scenarios.

[0004] In recent years, artificial intelligence technology has developed rapidly, especially with breakthroughs in image recognition, natural language processing, and 3D modeling, providing new approaches to optimizing traditional engineering problems. Deep learning networks (such as PointNet and PointNet++) that take point clouds as input have the ability to directly process unstructured 3D data and extract spatial feature information without relying entirely on regular meshes. Point cloud neural networks, by learning the intrinsic relationship between structural points and spatial responses, can establish an efficient and accurate mapping between 3D geometry and flow fields, greatly improving modeling efficiency and response prediction capabilities.

[0005] While point cloud networks have achieved significant results in fields such as 3D recognition and shape reconstruction, their application in the performance optimization of engineering machinery is still in its early stages of exploration. Especially in systems like centrifugal pumps, which involve complex internal flow fields, rotating components, and unsteady disturbances, a systematic optimization framework is lacking to effectively embed point cloud neural networks into the design process. Therefore, how to fully utilize the advantages of point cloud networks to construct optimization process models oriented towards engineering needs has become a current research hotspot and technical challenge in this field. Summary of the Invention

[0006] The purpose of this invention is to provide a centrifugal pump pressure pulsation optimization method and system based on point cloud neural networks, which realizes efficient and intelligent optimization between the internal structural parameters of the centrifugal pump and the pressure pulsation response, and quickly and accurately captures the sensitivity of complex geometric parameters to pressure pulsation, thereby achieving a significant reduction in pressure pulsation.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A centrifugal pump pressure pulsation optimization method based on point cloud neural networks includes the following steps:

[0009] Step 1: Collect characteristic structural parameters and rated operating condition performance parameters of centrifugal pumps to build a centrifugal pump structure-operating condition database;

[0010] Step 2: Based on the constructed database, perform CFD modeling and simulation, set monitoring points, collect performance parameters and pressure pulsation data, and generate centrifugal pump structure point cloud and flow field point cloud datasets and perform standardization processing.

[0011] Step 3: Add a structural parameter branch to the PointNet segmentation architecture. This branch uses a structural attention module to embed and encode the structural parameters and assign attention weights to extract structural feature representations. At the same time, the main network extracts the geometric-physical coupling features of the structural point cloud and the flow point cloud. The weighted structural attention features are concatenated with the point cloud features to form an interpretable structural-flow coupling composite feature vector, which is input into the MLP network to predict the pressure pulsation response value, thus constructing a point cloud neural network prediction architecture that integrates the structural parameter attention mechanism.

[0012] Step 4: Preprocess the standardized dataset from Step 2 into training and test sets. Input the training set into the neural network prediction architecture constructed in Step 3. Use end-to-end supervised learning to jointly optimize the hyperparameters of the structural parameter attention module and the point cloud feature extraction module to obtain a stress pulsation prediction model based on the structural parameter attention mechanism.

[0013] Step 5: Extract the structural parameters and attention weights of the trained model, combine them with the pulsation prediction values ​​of different structural samples, analyze the attention distribution and prediction variability, and construct a structural parameter sensitivity ranking matrix.

[0014] Step 6: Select the set of structural parameters with the highest attention weight or the highest sensitivity as optimization variables. Under the premise of ensuring that the performance parameters do not decrease, introduce an optimization strategy with "reducing pressure pulsation at key monitoring points" as the objective function. Combine the model prediction value as the objective feedback, and use the gradient descent method to perform multiple rounds of parameter combination iterations to search for the optimal structural parameters that minimize pressure pulsation.

[0015] Step 7: Perform closed-loop verification and update. Input the optimized parameters into the CFD platform for verification. If the pressure pulsation is significantly reduced and the head efficiency does not decrease, output the final centrifugal pump design scheme. Otherwise, add the verification data to the training set, fine-tune the model, and re-optimize.

[0016] Furthermore, the performance parameters of the rated operating condition in step 1 are: the rated operating condition is the rated flow rate and speed under the design parameters; the performance parameters include head and efficiency.

[0017] Further, step 3 includes:

[0018] Step 3.1: Perform high-dimensional embedding mapping on the structural parameter vector s to obtain the embedded feature vector e of the structural parameters;

[0019] Step 3.2: Calculate the weight vector α = Softmax(W2·ReLU(W1·s)) between the structural parameter dimensions using the self-attention mechanism; where W1 is the first fully connected layer weight matrix, mapping the input to the hidden layer dimension; W2 is the second fully connected layer weight matrix, with the output having the same number of structural parameters as the original; and α has the same dimension as s.

[0020] Step 3.3: Merge the attention weights and embedded features channel-wise using a weighted fusion method. k =α⊙e, where e k The weighted structural embedding features are represented by ⊙, which indicates element-wise multiplication.

[0021] Step 3.4: Place e k The geometric-flow coupling features extracted from the point cloud branches are fused and spliced ​​into a composite feature vector, which serves as the input for the subsequent pressure pulsation prediction model.

[0022] Furthermore, the generation of the composite feature vector satisfies:

[0023] f pointNet =concat(f Structural ,f Flow )

[0024] f used =concat(f pointNet ,e k )

[0025] Among them, f Structural The extracted structural point cloud feature vector, f Flow Extracted fluid point cloud feature vector, f pointNet The point cloud feature vectors extracted by the point cloud neural network, concat() represents the concatenation operation, merging the first and last ends of two vectors into a longer composite vector, f used The final composite feature vector is used to feed the pressure pulsation value into the prediction module MLP.

[0026] Furthermore, the standardized dataset in step 4 undergoes preprocessing, including:

[0027] The structural parameter vectors are directly normalized to form a fixed-length input format; the CSV files exported from CFD calculations are converted into binary PKL files; the point cloud data is divided into training and testing sets; and rotation and translation data augmentation methods are used to expand the training set.

[0028] Furthermore, in step 4, the mean squared absolute error (MAE) is selected as the loss function, and the Adam gradient optimization method is used to optimize the parameters. The preprocessed training set is imported into the point cloud neural network prediction architecture with fused structural parameter attention mechanism in a supervised learning manner. Hyperparameters are set until the model training converges, and the optimal weight information is saved.

[0029] The hyperparameter adjustment order is as follows: embedding layer dimension, attention layer size, number of MLP feature extraction layers, Adam optimizer learning rate, Dropout ratio, batch point cloud quantity, and training iteration cycle.

[0030] Further, step 5 includes:

[0031] Step 5.1: Using a self-attention mechanism, assign corresponding weights to each input structural parameter to reflect the contribution of the structural parameter to the prediction of pressure pulsations;

[0032] Step 5.2: Quantify the impact fluctuation of each structural parameter on the prediction result by calculating the variance or standard deviation of the pressure pulsation prediction result after fine-tuning the structural parameters;

[0033] Step 5.3: Combine attention weights and prediction variability to construct a sensitivity ranking matrix, where the value of each structural parameter represents its sensitivity score to changes in pressure pulsation.

[0034] Furthermore, the optimization objective function in step 6 is:

[0035]

[0036] Where s is the structural parameter vector, and x is the point cloud geometric input. H represents the predicted pressure fluctuations, and H and η represent the predicted head and efficiency of the network. base and η base The design head and efficiency under standard operating conditions based on design parameters; L target The minimum predicted pressure fluctuation value under the objective function constraint;

[0037] The gradient descent method:

[0038]

[0039] Where λ is the iteration rate, s t For the t-th iteration, To obtain the partial derivative.

[0040] Furthermore, the head and efficiency in step 7 are the monitoring parameters in the CFD solver.

[0041] Head calculation formula:

[0042] Where ρ is the density of water and g is the acceleration due to gravity;

[0043] The outlet pressure is the average total pressure of the point cloud at the outlet section.

[0044] Efficiency calculation formula:

[0045] Among them, Q v For export flow, P 轴 This refers to shaft power.

[0046] Furthermore, during the closed-loop verification update in step 7, when the CFD verification error exceeds the threshold, the verification data is added to the training set to fine-tune the model, and then the optimization steps of steps 5 to 6 are re-executed in the local neighborhood based on the fine-tuned model.

[0047] A computer device / apparatus / system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a centrifugal pump pressure pulsation optimization method based on a point cloud neural network.

[0048] The beneficial effects of this invention are as follows:

[0049] This invention proposes a centrifugal pump modeling and optimization strategy based on point cloud neural networks. Using centrifugal pump structural parameters as excitation, a customized point cloud neural network as the transfer function, and pressure pulsation changes as the response, a parameter optimization model is constructed. This method features improved computational efficiency, high computational accuracy, fast response speed, and strong optimization capabilities. Compared with traditional parameter traversal methods, it significantly reduces the time required for the same accuracy requirements, improving optimization computational efficiency several times over. It is particularly suitable for complex optimization scenarios under multiple performance constraints, as well as optimization tasks under complex structures and unsteady flow conditions. Simultaneously, it significantly improves the prediction accuracy and parameter control capability of centrifugal pump pressure pulsation response. It can not only capture the nonlinear relationship between complex flow and structural parameters, but also achieve the goal of minimizing pressure pulsation while meeting multiple performance constraints such as head and efficiency. Furthermore, the model structure is flexible and can be extended to multi-objective optimization problems, further supporting the coupled optimization needs of multiple performance aspects such as head, efficiency, noise, and vibration. Furthermore, the proposed parameter sensitivity ranking mechanism can effectively identify the key geometric factors that most significantly affect pressure pulsation, helping engineers to accurately tune parameters, reduce invalid design attempts, and assist designers in multi-objective parameter optimization in the early stages of structural design. This significantly shortens the design cycle and improves the intelligent optimization level of pump structures. This method exhibits good robustness and scalability, applicable to structural optimization tasks of pumps of different scales and types. Based on a more efficient and intelligent data-driven approach, it not only enhances the intelligence level of pump structural design but also provides a novel solution for fluid machinery performance optimization. It offers strong technical support for the intelligent design and manufacturing of centrifugal pumps and similar impeller machinery, demonstrating broad engineering application prospects and good sustainability. Attached Figure Description

[0050] Figure 1 This is a flowchart of the centrifugal pump pressure pulsation optimization of the present invention.

[0051] Figure 2 This is an example framework of the centrifugal pump point cloud neural network of the present invention.

[0052] Figure 3 These are the key conclusions and effect diagrams of the embodiments of the present invention. Detailed Implementation

[0053] The present invention will now be further described with reference to the accompanying drawings.

[0054] 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.

[0055] Please see Figures 1-2 This invention discloses a centrifugal pump pressure pulsation optimization method based on point cloud neural networks, comprising the following steps:

[0056] Step 1: Establish a centrifugal pump structure-operating condition database;

[0057] Key structural parameters of various centrifugal pumps, along with their corresponding rated operating conditions and performance parameters, were collected to establish a structural parameter table. This provides a prerequisite for subsequent CFD simulation and pressure pulsation optimization. The rated operating conditions refer to the rated flow rate and speed under the design parameters, while the performance parameters refer to the head and efficiency. Specific steps include:

[0058] (1) Collect ISO standard designs for various types of centrifugal pumps, such as common types of pumps like single-stage centrifugal pumps, multi-stage centrifugal pumps, mixed-flow pumps, and axial-flow pumps. The target type of pump to be optimized needs to be included in the database categories to ensure accuracy.

[0059] (2) Structural parameters are important modeling features of centrifugal pumps, such as blade inlet angle, blade outlet angle, impeller outlet diameter, impeller outlet width, number of blades, blade wrap angle, impeller clearance, blade thickness and other main structural features that contact flow and have an impact.

[0060] (3) Expand the design parameters of the centrifugal pump into several groups based on the design manual, national standards and other characteristic design ranges, and establish a structural parameter table.

[0061] Step 2: Collect point cloud data of the centrifugal pump structure and flow field;

[0062] Based on a CFD simulation platform, transient calculations were performed under standard operating conditions for multiple combinations of structural parameters. This involved extracting the structural point cloud of the centrifugal pump and defining key monitoring points in the flow field. These monitoring points were a set of coordinates artificially designated in the CFD simulation settings at the impeller and volute flow channels to record transient pressure values ​​at each time step. Subsequently, instantaneous pressure data from these monitoring points at different times were collected, mapped to a flow response point cloud, and pressure pulsations under the corresponding parameters were calculated. Finally, all point cloud data were standardized to a standard input format to establish a coupled point cloud dataset of structural geometry and flow behavior. Specific steps included:

[0063] (1) Simulate the centrifugal pump based on the CFD (Computational Fluid Dynamics) simulation platform, select multiple structural parameter combinations and perform transient calculations under standard operating conditions. The simulation model should include the internal geometry of the centrifugal pump and the dynamic behavior of its flow field.

[0064] (2) Extract the centrifugal pump structure point cloud data, flow field point cloud data, and monitoring point data from the CFD simulation results, and calculate the pressure pulsation value of the entire flow field at two different times.

[0065] (3) Standardize and unify the data format. All collected point cloud data should follow a unified standard format to ensure that the data of structural point clouds and flow behavior point clouds can be effectively connected and fused. Then, store the point cloud data in a fixed format, which can be used for subsequent training and prediction of point cloud neural network models.

[0066] Furthermore, the simulation described in section 2.1 is a transient calculation based on steady-state initial values, which facilitates the calculation of the periodic fluctuation values ​​of the pressure pulsation field.

[0067] Furthermore, the pressure pulsation value mentioned in 2.2 is the instantaneous pressure difference between two moments.

[0068]

[0069] Furthermore, the structural geometry dataset in section 2.2 includes structural spatial coordinates (x, y, z), operating condition labels, and structural parameter vectors; the flow field dataset refers to the flow field spatial coordinates (x, y, z) and the pressure fluctuation characteristics between two time points, as well as the predicted performance parameters.

[0070] Furthermore, the data unification format and standardization in section 2.3 refers to common data standardization methods such as data denoising, invalid point removal, and coordinate normalization of the original point cloud.

[0071] Step 3: Construct a point cloud neural network prediction architecture that integrates structural parameter attention mechanism;

[0072] This step adds a structural parameter attention branch to the traditional PointNet segmentation framework, and extracts features in parallel with the "structural point cloud branch" and "flow field point cloud branch". After completing the two-level feature fusion, the pressure pulsation prediction value is output.

[0073] Specifically, the structural parameters in the structural parameter attention branch are vectors, while the point cloud is high-dimensional geometric data. The branch's role is to "assign importance weights" to the structural parameters before they enter the network, enabling the network to automatically identify which structural parameters have a greater impact on pressure pulsation prediction. It acts as a "self-filter" and "feature amplifier" in the entire network.

[0074] Specifically, the structural parameter vector s is sequentially activated by a first linear layer (ReLU) and then by a second linear layer to obtain scores for each parameter. A softmax function is applied to each score to obtain a weight α (ranging from 0 to 1, with a total weight of 1), representing the importance percentage of each parameter under the current operating condition. Subsequently, the structural embedding features are multiplied channel-by-channel by α to form a weighted feature e. kThis amplifies key parameters and suppresses secondary parameters. The aforementioned weights are learned end-to-end with the pressure pulsation prediction loss, requiring no manual setting, and are used for subsequent fusion and prediction with point cloud features.

[0075] The attention allocation mechanism:

[0076] α = Softmax(W2·ReLU(W1·s))

[0077] Where s is the input structural parameter vector, W1 is the first fully connected layer weight matrix, which maps the input to the hidden layer dimension; W2 is the second fully connected layer weight matrix, and the output has the same number of structural parameters as the original; α is the output structural parameter attention weight vector, with the same dimension as s.

[0078] Furthermore, the characteristics of the weighted structure are:

[0079] e k =α⊙e

[0080] Where e is the embedded feature vector of the structural parameters, α is the attention weight vector, and ⊙ represents element-wise multiplication. k This is a weighted structural embedding feature, resulting in a vector that emphasizes important structural information.

[0081] Specifically, a "structural parameter attention" branch was added to the original point cloud network, working simultaneously with the "structural point cloud branch" and the "flow field point cloud branch." The former reads design parameters (such as the number of blades, outlet angle, diameter, and width) and automatically calculates a set of "weights" to tell the model which parameters are more critical. Meanwhile, the structural point cloud branch learns the mechanical shape characteristics of the pump's 3D geometric point cloud, and the flow field point cloud branch learns flow characteristics from point clouds with pulsating pressure. Then, the "structural point cloud features" and "flow field point cloud features" are combined, and the "weighted structural parameter features" are incorporated as conditions to output pressure pulsation predictions for monitoring points. The entire process is end-to-end training, with weights automatically learned. This makes the model more accurate and also identifies which structural parameters have the greatest impact, facilitating subsequent optimization.

[0082] Furthermore, the feature vector fusion and concatenation mechanism:

[0083] f pointNet =concat(f Structural ,f Flow )

[0084] f used =concat(f pointNet ,e k )

[0085] Among them, f Structural The extracted structural point cloud feature vector, fFlow Extracted fluid point cloud feature vector, f pointNet The point cloud feature vectors extracted by the point cloud neural network, concat() represents the concatenation operation, merging the first and last ends of two vectors into a longer composite vector, f used The final composite feature vector is used to feed the pressure pulsation value into the prediction module MLP.

[0086] Step 4: Train a stress pulsation prediction model with an attention mechanism;

[0087] The mean squared absolute error (MAE) is selected as the loss function, and the Adam gradient optimization method is used to optimize the parameters. The preprocessed training set is imported into the point cloud neural network prediction architecture with fused structural parameter attention mechanism in a supervised learning manner. Appropriate hyperparameters are set until the model training converges, and the optimal weight information is saved.

[0088] Specifically, the hyperparameters are adjusted in the following order: embedding layer dimension, attention layer size, number of MLP feature extraction layers, Adam optimizer learning rate, Dropout ratio, batch point cloud size, and training iteration cycle.

[0089] Specifically, the number of MLP feature extraction layers needs to be considered comprehensively based on the complexity of the point cloud structure features, computing resources, and the desired prediction accuracy.

[0090] Specifically, the initial learning rate of the Adam optimizer needs to be dynamically adjusted based on the training results. Setting an excessively large learning rate can cause the model to miss the optimal solution, leading to overfitting; setting an excessively small learning rate will slow down the learning process. The learning rate needs to be set based on experience and continuous experimentation.

[0091] Specifically, the Dropout ratio is used to prevent overfitting during model training by randomly "dropping" some input units (i.e., neurons) and setting their values ​​to zero. This avoids the network's over-reliance on certain neurons, prompting the network to learn redundant features, thereby improving the model's generalization ability. It needs to be tested comprehensively based on the amount of training data and the number of neural network layers and units; an empirical value is 0.1 to 0.5.

[0092] Specifically, the batch point cloud size is the number of samples selected from the training set in each training iteration, which is determined based on computer performance;

[0093] Specifically, the training iteration cycle refers to the number of times the entire training set is input to ensure that the model can learn the patterns in the data through a sufficient number of training processes, thereby achieving a better fitting effect.

[0094] Specifically, supervised learning refers to the acceptable level of mean squared absolute error (MAE). When it is sufficiently small, it is considered that the optimal learning weights have been found.

[0095] Specifically, deep learning algorithms must be performed on high-performance computers. The entire neural network training model is based on the PyTorch platform, utilizing the high bandwidth and multi-threaded parallel computing capabilities of GPUs for training.

[0096] Step 5: Perform parameter sensitivity ranking analysis based on attention output;

[0097] After model training, the weight outputs of the structural parameter attention layer are extracted as a quantitative indicator of the influence of each structural parameter on pulsation prediction. By combining the pulsation prediction values ​​of different structural samples and comprehensively analyzing the attention distribution and prediction variability, a structural parameter sensitivity ranking matrix is ​​constructed. This matrix can be used to identify which structural parameters are most sensitive to changes in target pressure pulsation, thereby guiding the direction of structural optimization.

[0098] Specifically, the attention weight allocation is performed by forward reasoning on the test set after fixing the parameters of the trained network. The structural parameter vector s is passed through a fully connected network and the ReLU activation function in sequence, then through another fully connected network and subjected to Softmax normalization. Finally, the output of the structural parameter attention branch is taken as the importance weight of each structural parameter.

[0099] Specifically, without violating engineering constraints (head and efficiency not lower than their original values), only one structural parameter is modified at a time, making small adjustments upwards and downwards within the allowable design range. After each adjustment, the trained network is asked to re-predict the pressure pulsations at the monitoring point. If a slight adjustment causes a large change in the prediction result, it indicates that the parameter is "very sensitive" to pulsations; a small change indicates it is "not so sensitive." Probes that do not meet the head / efficiency constraints are discarded or downweighted. This process is repeated across multiple samples, multiple operating conditions, and multiple monitoring points to obtain a stable and reliable measure of "how much variation there is."

[0100] Specifically, by analyzing the variability of the structural parameters attention weights and pressure pulsation prediction results obtained during training, a total score is assigned to each structural parameter, and a "sensitivity ranking" matrix is ​​listed for each monitoring point. This matrix includes the degree of influence of different structural parameters on pressure pulsations.

[0101] Furthermore, the degree of influence of pressure pulsation is obtained through the following steps:

[0102] (1) Attention weight allocation: Through the self-attention mechanism, a corresponding weight is assigned to each input structural parameter (such as the number of blades Z, the outlet angle β2, etc.). The weight reflects the contribution of the structural parameter to the predicted pressure pulsation.

[0103] (2) Prediction variability analysis: By calculating the variance or standard deviation of the pressure pulsation prediction results after the structural parameters are fine-tuned, the influence of each structural parameter on the prediction results is quantified.

[0104] (3) Sensitivity Ranking: A sensitivity ranking matrix is ​​constructed by combining attention weights and prediction variability, where the value of each structural parameter represents its sensitivity to pressure pulsation changes. This matrix provides a clear reference for subsequent optimization design, prioritizing the adjustment of structural parameters that have the greatest impact on pressure pulsation.

[0105] Step Six: Optimize the target-driven structural parameters based on attention ranking;

[0106] After obtaining the "parameter sensitivity ranking matrix" in step five, parameters with high influence on the target monitoring point are selected as optimization variables. Without reducing the head and efficiency, the structural combination that minimizes the pressure pulsation at the key monitoring point is iteratively searched.

[0107] Specifically, several structural parameters with high ranking and influence are selected as optimization variables, and a trained three-branch point cloud neural network is used as a fast prediction model. First, the feasible range and step size are determined for each variable (the number of blades is taken as an integer, and the outlet angle, diameter, and width are taken as continuous intervals). Using the existing benchmark scheme as the initial solution, an optimization objective is established: "to minimize the pressure pulsation at the target monitoring point while ensuring that the head and efficiency are not lower than the initial values."

[0108] Specifically, the system probes variables in descending order of sensitivity, making small increases or decreases to individual variables, and then uses the predictive model to calculate pressure pulsation, head, and efficiency in real time. An update is only accepted as a valid improvement if both head and efficiency simultaneously meet the baseline constraints; probes that do not meet the constraints are discarded.

[0109] The expressions for the objective function and constraints are as follows:

[0110]

[0111] Where s is the structural parameter vector, and x is the point cloud geometric input. H represents the predicted pressure fluctuations, and H and η represent the head and efficiency predicted by the network. base and η base The design head and efficiency under standard operating conditions based on design parameters; L target The minimum predicted pressure fluctuation value under the objective function constraint;

[0112] Specifically, the gradient descent method employs an adaptive step size strategy. If the current attempt significantly reduces the target value, the step size is appropriately increased to accelerate convergence; if there is no improvement, the step size is halved to refine the search. If a variable update goes out of bounds, it is immediately projected back to the feasible region boundary to ensure that all candidates are within the engineering allowable range.

[0113] The gradient descent method:

[0114]

[0115] Where λ is the iteration rate, s t For the t-th iteration, To obtain the partial derivative;

[0116] Furthermore, to avoid local optima, the system employs parallel search with multiple starting points or multiple sets of high-impact parameters, and dynamically updates the "set of variables participating in optimization" based on the attention weight of the current sample after several rounds of search, so that computing resources are continuously focused on the most sensitive parameter dimensions, thereby improving search efficiency and stability.

[0117] Furthermore, the process terminates when the target improvement falls below a threshold or the maximum number of iterations is reached after several consecutive rounds, outputting the optimal combination of structural parameters and the corresponding predicted pressure pulsation, head, and efficiency. A CFD verification is performed on the optimal solution; if the verification and prediction deviation exceeds the tolerance, the verification sample is used for small-batch fine-tuning, and the search continues in the neighborhood of this solution until the tolerance is met.

[0118] Step 7: Output optimization parameters and perform CFD verification and structural closed-loop update;

[0119] The optimized structural parameter combination is input into a CFD platform for secondary verification, simulating its flow field and pressure pulsation response to verify the prediction accuracy and engineering applicability of the optimized model. If the optimization effect is significant without sacrificing head and efficiency, the structure is output as the final centrifugal pump design scheme, achieving accurate mapping and closed-loop design between structural parameters and pressure pulsation, and completing the development of a high-performance, low-pulsation pump product.

[0120] Specifically, based on the optimal combination of structural parameters obtained in step six, a secondary verification is performed in the CFD platform using the same method. The optimized scheme is then subjected to transient simulation under the design conditions and several deviated conditions. Pressure pulsation values ​​at the same locations in the dataset are extracted, and the head and efficiency are output simultaneously.

[0121] Furthermore, head and efficiency are monitored parameters in the CFD solver. The head calculation formula is as follows:

[0122]

[0123] Where ρ is the density of water, and g is the acceleration due to gravity.

[0124] The outlet pressure is the average total pressure of the point cloud at the outlet section.

[0125] Efficiency calculation formula:

[0126]

[0127] Among them, Q v For export flow, P 轴 This refers to shaft power.

[0128] Specifically, the CFD results are compared with the model predictions from step six, and acceptance is performed according to preset tolerance criteria. The average relative error of the pulsation index at key monitoring points is not higher than a predetermined threshold (preferably ≤10%), and the head and efficiency do not decrease compared to the initial values. When all indicators meet the criteria, the optimization scheme is confirmed to be effective, and the structural parameters are output as the final design scheme. A complete set of technical documents, including parameter tables, CFD verification results, and operating condition boundaries, is then generated for detailed design or manufacturing.

[0129] Furthermore, if any key indicator fails to meet the standard, a closed-loop update is initiated. The CFD validation data (geometry, operating conditions, monitoring point pulsation, head, and efficiency) is added to the training set. Small-batch fine-tuning is performed as in step four to correct the model. Subsequently, following the procedures in steps five and six, a rapid search is conducted again in the current solution neighborhood on the corrected predictor until a new solution that meets the tolerance is obtained without sacrificing head and efficiency. The final output optimization report includes: parameter changes, pulsation reduction, head and efficiency maintenance, CFD comparison error, closed-loop convergence trajectory, and manufacturability conclusions, completing the engineering implementation of the high-efficiency, low-pulsation centrifugal pump.

[0130] In particular, in some preferred embodiments of the present invention, a computer device is also provided, including a memory and a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the centrifugal pump pressure pulsation optimization method based on point cloud neural network described in any of the above embodiments.

[0131] In some other preferred embodiments of the present invention, a computer-readable storage medium is also provided, on which a computer program / instruction is stored, wherein when the computer program is executed by a processor, the steps of the centrifugal pump pressure pulsation optimization method based on point cloud neural network described in any of the above embodiments are implemented.

[0132] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above embodiments of the centrifugal pump pressure pulsation optimization method based on point cloud neural networks, which will not be repeated here.

[0133] Computer-readable storage media encompass a variety of types, including persistent and non-persistent, portable and fixed. These media store information using different technologies, and the content can be machine instructions, data structures, program modules, or other types of data. Some typical examples of computer storage media include: phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), various types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory and other storage technologies, optical storage media such as CD-ROM and digital video disc (DVD), magnetic storage devices such as magnetic tape and disks, and other non-transferable media used to store information accessible to computing devices. It is important to note that the computer-readable media described herein do not include temporary storage media, such as modulated data signals and carrier waves.

[0134] Those skilled in the art will further recognize that the operation of the module can be achieved using existing technical protocols or programs, without relying on new computer programs themselves. The units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0135] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0136] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing centrifugal pump pressure pulsation based on point cloud neural networks, characterized in that: Includes the following steps: Step 1: Collect characteristic structural parameters and rated operating condition performance parameters of centrifugal pumps, and construct a centrifugal pump structure-operating condition database; Step 2: Based on the constructed database, perform CFD modeling and simulation, set monitoring points, collect performance parameters and pressure pulsation data, and generate centrifugal pump structure point cloud and flow field point cloud datasets and perform standardization processing. Step 3: Add a structural parameter branch to the PointNet segmentation architecture. This branch uses a structural attention module to embed and encode the structural parameters and assign attention weights to extract structural feature representations. At the same time, the main network extracts the geometric-physical coupling features of the structural point cloud and the flow point cloud. The weighted structural attention features are concatenated with the point cloud features to form an interpretable structural-flow coupling composite feature vector, which is input into the MLP network to predict the pressure pulsation response value, thus constructing a point cloud neural network prediction architecture that integrates the structural parameter attention mechanism. Step 4: Preprocess the standardized dataset from Step 2 into training and test sets. Input the training set into the neural network prediction architecture constructed in Step 3. Use end-to-end supervised learning to jointly optimize the hyperparameters of the structural parameter attention module and the point cloud feature extraction module to obtain a stress pulsation prediction model based on the structural parameter attention mechanism. Step 5: Extract the structural parameters and attention weights of the trained model, combine them with the pulsation prediction values ​​of different structural samples, analyze the attention distribution and prediction variability, and construct a structural parameter sensitivity ranking matrix. Step 6: Select the set of structural parameters with the highest attention weight or the highest sensitivity as optimization variables. Under the premise of ensuring that the performance parameters do not decrease, introduce an optimization strategy with "reducing pressure pulsation at key monitoring points" as the objective function. Combine the model prediction value as the objective feedback, and use the gradient descent method to perform multiple rounds of parameter combination iterations to search for the optimal structural parameters that minimize pressure pulsation. Step 7: Perform closed-loop verification and update. Input the optimized parameters into the CFD platform for verification. If the pressure pulsation is significantly reduced and the head efficiency does not decrease, output the final centrifugal pump design scheme. Otherwise, add the verification data to the training set, fine-tune the model, and re-optimize.

2. The centrifugal pump pressure pulsation optimization method based on point cloud neural network according to claim 1, characterized in that: The performance parameters of the rated operating condition in step 1 are: rated operating condition is the rated flow rate and speed under the design parameters; performance parameters include head and efficiency.

3. The centrifugal pump pressure pulsation optimization method based on point cloud neural network according to claim 1, characterized in that: Step 3 includes: Step 3.1: Perform high-dimensional embedding mapping on the structural parameter vector s to obtain the embedded feature vector e of the structural parameters; Step 3.2: Calculate the weight vector α = Softmax(W2·ReLU(W1·s)) between the structural parameter dimensions using the self-attention mechanism; where W1 is the first fully connected layer weight matrix, mapping the input to the hidden layer dimension; W2 is the second fully connected layer weight matrix, with the output having the same number of structural parameters as the original; and α has the same dimension as s. Step 3.3: Merge the attention weights and embedded features channel-wise using a weighted fusion method. k =α⊙e, where e k The weighted structural embedding features are represented by ⊙, which indicates element-wise multiplication. Step 3.4: Place e k The geometric-flow coupling features extracted from the point cloud branches are fused and spliced ​​into a composite feature vector, which serves as the input for the subsequent pressure pulsation prediction model.

4. The centrifugal pump pressure pulsation optimization method based on point cloud neural network according to claim 3, characterized in that: The generation of the composite feature vector satisfies: f pointNet =concat(f Structural ,f Flow ) f used =concat(f pointNet ,e k ) Among them, f Structural The extracted structural point cloud feature vector, f Flow Extracted fluid point cloud feature vector, f pointNet The point cloud feature vectors extracted by the point cloud neural network, concat() represents the concatenation operation, merging the first and last ends of two vectors into a longer composite vector, f used The final composite feature vector is used to feed the pressure pulsation value into the prediction module MLP.

5. The centrifugal pump pressure pulsation optimization method based on point cloud neural network according to claim 1, characterized in that: In step 4, mean squared absolute error (MAE) is selected as the loss function, and the Adam gradient optimization method is used to optimize the parameters. The pre-processed training set is imported into the point cloud neural network prediction architecture with fused structural parameter attention mechanism in a supervised learning manner. Hyperparameters are set until the model training converges and the optimal weight information is saved. The hyperparameter adjustment order is as follows: embedding layer dimension, attention layer size, number of MLP feature extraction layers, Adam optimizer learning rate, Dropout ratio, batch point cloud quantity, and training iteration cycle.

6. The centrifugal pump pressure pulsation optimization method based on point cloud neural network according to claim 1, characterized in that: Step 5 includes: Step 5.1: Using a self-attention mechanism, assign corresponding weights to each input structural parameter to reflect the contribution of the structural parameter to the prediction of pressure pulsations; Step 5.2: Quantify the impact fluctuation of each structural parameter on the prediction result by calculating the variance or standard deviation of the pressure pulsation prediction result after fine-tuning the structural parameters; Step 5.3: Combine attention weights and prediction variability to construct a sensitivity ranking matrix, where the value of each structural parameter represents its sensitivity score to changes in pressure pulsation.

7. The centrifugal pump pressure pulsation optimization method based on point cloud neural network according to claim 1, characterized in that: The optimization objective function in step 6 is: Where s is the structural parameter vector, and x is the point cloud geometric input. H represents the predicted pressure fluctuations, and H and η represent the predicted head and efficiency of the network. base and η base The design head and efficiency under standard operating conditions based on design parameters; L target The minimum predicted pressure fluctuation value under the objective function constraint; The gradient descent method: Where λ is the iteration rate, s t For the t-th iteration, To obtain the partial derivative.

8. The centrifugal pump pressure pulsation optimization method based on point cloud neural network according to claim 1, characterized in that: The head and efficiency in step 7 are the monitoring parameters in the CFD solver. Head calculation formula: Where ρ is the density of water and g is the acceleration due to gravity; The outlet pressure is the average total pressure of the point cloud at the outlet section. Efficiency calculation formula: Among them, Q v For export flow, P 轴 This refers to shaft power.

9. The centrifugal pump pressure pulsation optimization method based on point cloud neural network according to claim 1, characterized in that: When performing closed-loop verification update in step 7, if the CFD verification error exceeds the threshold, the verification data is added to the training set to fine-tune the model, and then the optimization steps of steps 5 to 6 are re-executed in the local neighborhood based on the fine-tuned model.

10. A computer device / equipment / system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 9.

Images