A method and system for predicting particle settling and collision and flow characteristics in a pump

Abstract

The application provides a method and system for predicting particle settling, collision and flow characteristics in a pump, comprising: constructing a multi-phase flow training sample library of real-time collected environment and virtual-real fusion; after pre-processing the sample library data, using a double-flow deep neural network architecture to extract multiple source feature vectors in parallel; cross-fusing the multiple source feature vectors to obtain a multi-modal fusion feature vector, and establishing an Euler-Lagrange two-way coupled partial differential control equation; inputting the multi-modal fusion feature vector into a physical information neural network based on a gradient-aware adaptive physical loss function, and outputting particle instantaneous settling velocity, particle collision contact force, local entropy production rate and global reconstructed three-dimensional flow field; based on the output calculation index, hierarchical regulation is executed accordingly. The application breaks the feature mismatch problem, replaces the complex grid iteration with millisecond-level forward reasoning conforming to the conservation law of fluid mechanics, and realizes high-precision intelligent online prediction of complex flow field characteristics.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of fluid mechanics, numerical simulation of solid-liquid two-phase flow and artificial intelligence, specifically to a method and system for predicting particle settling, collision and flow characteristics in a pump. Background Technology

[0002] In scenarios such as slurry transportation, dredging projects, deep-sea mining, and sand-containing media transport, pumps, as the core power equipment for transporting solid-liquid two-phase media, directly affect transport efficiency, operational stability, and service life due to their internal flow state. Unlike single-phase flow, particulate media flowing within a pump are affected not only by mainstream transport, pressure gradients, and turbulent pulsations, but also by a combination of factors including particle gravity settling, centrifugal migration, inter-particle collisions, and particle-wall collisions. This can easily lead to problems such as local enrichment, enhanced energy dissipation, intensified wall erosion, and increased risk of blockage in the impeller inlet area, low-speed region of the impeller passage, impeller outlet area, and near the volute tongue, resulting in decreased head, reduced efficiency, and intensified pressure pulsations. Currently, the CFD-DEM method is commonly used to simulate particle motion within the pump and its impact on flow characteristics. While this method can describe particle trajectories and two-phase coupling processes, processes such as particle drag, lift, collision recovery, and wall rebound largely rely on empirical formulas or semi-empirical models, exhibiting strong parameter dependence and poor adaptability across operating conditions. Meanwhile, the sedimentation and collision behavior of particles inside the pump has strong nonlinear and multi-scale coupling characteristics. Traditional empirical models are difficult to accurately characterize its complex dynamic process and have high computational costs, making it difficult to meet the needs of rapid prediction and intelligent optimization. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for predicting particle settling, collision, and flow characteristics within a pump.

[0004] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0005] A method for predicting particle settling, collision, and flow characteristics within a pump includes the following steps:

[0006] Step S1: Construct a training sample library for multiphase flow that integrates real-time acquisition environment with virtual-real fusion, and acquire flow field image data, particle trajectory data and simulation label data;

[0007] Step S2: Preprocess the flow field image data, particle trajectory data and simulation label data obtained in Step S1 to obtain flow field tensor features and particle topological features. Use a dual-flow deep neural network architecture to extract the features in parallel to obtain multi-source feature vectors. The multi-source feature vectors include flow field microstructure feature vectors and particle interaction feature vectors.

[0008] Step S3: Cross-fuse the multi-source feature vectors obtained in step S2 to obtain multi-modal fused feature vectors, and establish the Euler-Lagrange bidirectional coupled partial differential control equations;

[0009] Step S4: Input the multimodal fusion feature vector obtained in step S3 into the physical information neural network model based on gradient-aware adaptive physical loss function, and output the instantaneous particle settling velocity, particle collision contact force, local entropy yield and global reconstructed three-dimensional flow field.

[0010] Step S5: Calculate the pump performance factor, particle settling risk, particle erosion wear risk, and flow field stability index based on the output results of step S4, and perform graded adjustment accordingly.

[0011] Furthermore, the dual-flow deep neural network architecture includes a flow field perception branch and a particle interaction branch.

[0012] Furthermore, the flow field sensing branch structure is as follows:

[0013] Input layer: Receives preprocessed local flow field slice tensors centered on the target particle;

[0014] Convolutional layers: Based on a custom ResNet variant structure, it contains 4 residual modules. Each residual module contains 2 convolutional layers and a batch normalization layer. The first two residual modules have 32 output channels, the third residual module has 64 output channels, and the fourth residual module has 128 output channels.

[0015] Output layer: After global average pooling, a 128-dimensional flow field microstructure feature vector is output, which is used to characterize the local shear rate and vortex intensity of the tip leakage flow.

[0016] Furthermore, the particle interaction branching structure is as follows:

[0017] Input layer: Receives data centered on the target particle with a radius R = 15d. p The constructed local particle topology map data, where d p The diameter of the particle;

[0018] Graph Attention Layer: There are 2 layers in total. The first layer uses a multi-head attention mechanism to extract high-order collision association features of neighbor nodes; the second layer uses a single-head attention mechanism for feature aggregation and dimensionality reduction.

[0019] Output layer: The Sum-Pooling operation is used to output a 64-dimensional particle interaction feature vector, which is used to quantify the degree of particle aggregation and wake shielding effect in the pump.

[0020] Furthermore, the local flow field slice tensor incorporates the spatial distribution of the flow field Q criterion and the turbulent kinetic energy channels calculated by the physical feature constructor; the local particle topology graph data defines particles as nodes and the distance between particles as edges.

[0021] Furthermore, the process of generating the multimodal fusion feature vector is as follows: the weights of the contributions of the core vortex region and the high shear region of the flow field to particle collision and sedimentation are automatically calculated by the adaptive attention module, thereby updating the weight allocation of the flow field microstructure feature vector and the particle interaction feature vector, and then the weighted multimodal fusion feature vector is generated by dynamic splicing.

[0022] Furthermore, the Euler-Lagrange bidirectional coupled partial differential governing equations include:

[0023] The continuity equation and the Navier-Stokes equation are as follows:

[0024]

[0025]

[0026] In the formula, It is a vector differential operator; For fluid velocity, For fluid pressure, For time, and These are fluid density and dynamic viscosity, respectively. This is the momentum source term for the particle group acting on the fluid;

[0027] The particle dynamics equation is as follows:

[0028]

[0029] In the formula, For particle mass, Instantaneous settlement velocity, For alternating traction, Net weight and buoyancy, For virtual mass force; This refers to the contact force during particle collisions.

[0030] Particle collision contact force The particle interaction feature vector output by the particle interaction branch is denoted as h. p The mapping yields, i.e.:

[0031]

[0032] In the formula, The network is a mapping network consisting of fully connected layers and nonlinear activation functions, with the following parameters: ;

[0033] and local irreversible entropy production The equation, serving as the physical constraint for energy dissipation, is expressed as follows:

[0034]

[0035] In the formula, Let be the viscous dissipation function. The energy dissipation rate caused by inelastic particle collisions; T is the thermodynamic temperature.

[0036] Furthermore, the specific structure of the physical information neural network model is as follows:

[0037] Input layer: Receives multimodal fused feature vectors and cascades granular instantaneous Reynolds numbers Re. p With relative slip velocity u slip ;

[0038] Fully connected layers: 4 layers in total. Layer 1 has 256 neurons with a dropout rate of 0.2; Layer 2 has 128 neurons with a dropout rate of 0.2; Layer 3 has 64 neurons with a dropout rate of 0.1; Layer 4 has 32 neurons with a dropout rate of 0. The activation function used for all layers is the second-order continuously differentiable Swish function.

[0039] Output layer: Outputs predicted instantaneous particle settling velocity, particle collision contact force field, local entropy yield, and globally reconstructed 3D flow field;

[0040] Model Training and Physical Constraints: After each forward propagation, the automatic differentiation interface is called to construct the residuals of the fluid continuity equation, the Navier-Stokes equation, the particle dynamics equation, and the entropy yield equation as an energy constraint. The optimizer employs a hybrid optimization algorithm combining Adam and L-BFGS to construct a gradient-aware adaptive physical loss function. The formula is:

[0041]

[0042] in, , , , , , All are coefficients. The data error between the model predictions and the direct numerical simulation of particle analysis; , These are the residuals of the fluid continuity equation and the Navier-Stokes equation, respectively. The residuals of the particle dynamics equations substituted into the model predictions; The residual of the entropy productivity equation; For the particle dynamics topological constraint residuals in the particle interaction branch; For a dynamic adaptive weight tensor, x represents the spatial coordinates of the flow field, and t represents time.

[0043] Furthermore, the specific logic of the hierarchical adjustment is as follows:

[0044] Level 1 performance compensation: When the pump performance factor When the value is below 0.90, it is determined that the hydraulic performance has deteriorated or the output power is insufficient. At this time, the frequency converter is driven to increase the speed to compensate for the head or flow loss.

[0045] Secondary sedimentation intervention: when the particle settling risk R Settle When the value exceeds 1.00, it is determined that the local flow field is insufficient to effectively carry particles. At this time, the guide vane angle is adjusted to improve the suspension state of particles by optimizing the internal flow channel structure and prevent accumulation and blockage.

[0046] Level 3 wear protection: When the risk of particle erosion wear is R wear When the value exceeds 1.20, it is determined that the particle kinetic energy is too large, causing serious damage to the flow components. At this time, the rotation speed should be adjusted to reduce the impact intensity of the particles on the wall and extend the equipment life.

[0047] Level 4 Instability Warning: When the smooth stability S flow When the value exceeds 0.30, it is determined that the system has entered the unsteady operation zone. At this time, the combined speed and guide vane adjustment program is triggered to forcibly rebuild the stable flow field and send an abnormal operation warning signal to the central control room.

[0048] A system for predicting particle settling, collision, and flow characteristics within a pump, used to execute the method for predicting particle settling, collision, and flow characteristics within a pump, comprising:

[0049] A multi-source heterogeneous data acquisition module is used to acquire real-time video streams of high-speed flow fields and particle motion trajectory signals inside the pump, and integrates a high-fidelity numerical simulation data interface.

[0050] The multi-scale feature extraction module is used to extract the microstructure features of the flow field and the topological interaction features of the particles in parallel using the flow field perception branch and the particle interaction branch, respectively, to generate a multimodal fusion feature vector.

[0051] The physical information fusion prediction module is used to input the feature vector into a pre-trained physical information neural network based on gradient perception and adaptive physical loss function, and predict instantaneous collision force and settling velocity, local entropy yield and global reconstruction of three-dimensional flow field under the physical constraints of the Navier-Stokes equations and dynamic laws.

[0052] The operational condition quantitative assessment module is used to construct the pump performance factor η and particle settling risk R based on the instantaneous particle settling velocity, particle collision contact force, local entropy yield, and globally reconstructed three-dimensional flow field output by the physical information fusion prediction module. Settle Risk of particle erosion and wear (R) wear and flow field stability S flow This allows for a quantitative assessment of the flow state within the pump, particle deposition trends, erosion and wear levels, and the risk of unsteady operation, thus quantifying the flow state and wear risk.

[0053] The intelligent feedback control module is used to execute graded alarms based on the evaluation results and to adjust equipment operating parameters such as pump speed and guide vane angle in conjunction with the results.

[0054] The beneficial effects of this invention are as follows:

[0055] (1) The present invention adopts a dual-branch architecture of flow field sensing branch and particle interaction branch, which fundamentally solves the cross-scale feature mismatch problem of two-phase flow continuous field-discrete phase; the flow field sensing branch accurately extracts the spatial structural features of microscale vortices and shear rate of fluid continuous phase, while the particle interaction branch constructs the particle discrete phase into a dynamic topology map to quantify collision energy interaction. The cross-fusion of the two breaks the feature fuzziness defect of single network and truly restores the micro-mechanical mechanism of flow field centrifugal capture of particles and wake interference.

[0056] (2) By introducing a physical conservation mechanism, this invention breaks through the high dependence of the traditional CFD-DEM method on macroscopic static empirical formulas; by embedding the NS equation, dynamic law and entropy production rate equation characterizing energy dissipation as physical constraints into the network, the black box data fitting is transformed into parameter inversion following fluid dynamics, which not only eliminates the cumulative error of the traditional collision model in the unsteady strong swirling region in the pump, but also endows the model with strong generalization ability under the deviation of the training set working conditions.

[0057] (3) This invention constructs a gradient-aware adaptive loss function, which overcomes the network convergence problem in complex flow field regions with high gradients; for the severe pressure gradient region in fluid machinery, an adaptive weight tensor positively correlated with the local velocity gradient is introduced, which drives the model to spontaneously focus on the core vortex region with high shear and high dissipation during training, effectively avoiding the defect of conventional PINN being prone to local optima, and significantly improving the fidelity of particle contact force prediction in drastic flow fields.

[0058] (4) This method realizes a leap in computing power from implicit iteration to pure forward propagation, which strongly supports millisecond-level intelligent closed-loop early warning. After the model is trained offline, online prediction only needs to execute the forward inference of the network, which completely avoids the complicated mesh mapping and equation iteration in traditional fluid-structure interaction, compresses the time-consuming transient flow field calculation to millisecond-level response, effectively supports the hierarchical online early warning based on instantaneous particle state and collision intensity, and greatly improves the predictive maintenance level of fluid machinery. Attached Figure Description

[0059] Figure 1 This is a flowchart of the method for predicting particle settling, collision, and flow characteristics in a pump as described in this invention.

[0060] Figure 2 This is a schematic diagram of the physical information neural network model based on gradient-aware adaptive physical loss function as described in this invention.

[0061] Figure 3 This is a logic diagram of the online evaluation and intelligent classification early warning of the pump flow characteristics described in this invention. Detailed Implementation

[0062] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0063] Example 1

[0064] like Figure 1 As shown, the specific implementation steps of the method for predicting particle settling, collision, and flow characteristics in a pump according to the present invention are as follows:

[0065] S1: Construct a training sample library for multiphase flow that integrates real-time acquisition environment with virtual-real fusion. The acquired raw data includes pump flow field images, particle trajectories, etc., and high-fidelity numerical simulation is used to generate labeled data that is spatiotemporally aligned with the raw data. The labeled data includes at least the true values ​​of instantaneous particle settling velocity, particle collision contact force, and entropy yield reference values ​​characterizing energy dissipation. This step aims to solve the problem of direct measurement of particle surface forces in traditional methods by adopting a strategy of providing input from physical experiments and providing true values ​​from virtual simulation. Specifically:

[0066] S1.1: A non-invasive optical acquisition unit is deployed on the outside of the transparent multi-stage mixed-transport pump test bench. This unit includes two high-resolution high-speed industrial cameras (sampling frequency set at 2000 frames / second, resolution of 2048×2048 pixels) arranged at a 90-degree angle and a matching macro lens. Together with a high-frequency continuous-wave laser sheet light source (wavelength 532nm) extended by a cylindrical mirror, it performs precise tomographic scanning of the highly turbulent shear region of the pump impeller leading edge and blade tip gap, acquiring images of the flow field inside the pump, particle trajectories, etc.

[0067] S1.2: Simultaneously, a virtual data domain was constructed using the granular analytical direct numerical simulation (PR-DNS) technique. The Navier-Stokes (NS) equations were solved at a grid resolution of tens of millions of pixels, simulating Reynolds numbers in the range of 10. 4 Up to 10 6 The turbulent flow field within the pump was analyzed. Dynamic parameters such as particle settling velocity and collision contact force were calculated by integration, and the collision energy dissipation value of particles in the tip leakage vortex was extracted. Among them, particle settling velocity and particle collision contact force were used as truth labels for data supervision terms, and collision energy dissipation value was used as a reference value for entropy production rate equation constraints or energy dissipation loss terms. Together, they were used to construct the loss function of the subsequent physical neural network, thereby generating a virtual-real fusion multiphase flow pre-training sample library.

[0068] S2: The flow field image data, particle trajectory data, and simulation label data obtained in S1 are denoised, registered, sliced, normalized, and graph structured to obtain flow field tensor features and particle topological features. Specifically, the image data undergoes denoising, calibration, cropping, and normalization; the trajectory data undergoes trajectory association, interpolation completion, and coordinate unification; the simulation label data undergoes time alignment, spatial registration, outlier removal, and dimensionless processing; and a graph structure is constructed based on the particle trajectory data for the target particle neighborhood. The flow field tensor features and particle topological features are extracted in parallel using a two-stream deep neural network architecture, as detailed below:

[0069] S2.1: Flow Field Awareness Branch (CNN Network Structure): The input layer receives a preprocessed local flow field slice tensor centered on the target particle. This tensor incorporates the spatial distribution of the flow field Q-criterion and the turbulent kinetic energy (TKE) channel calculated by the physical feature constructor. The convolutional layer is based on a customized ResNet variant structure, containing four residual modules. Each residual module contains two convolutional layers and a batch normalization layer. The first two residual modules have 32 output channels, the third has 64, and the fourth has 128. The number of channels refers to the number of channels in the feature maps output by the convolutional layers in each residual module. The output layer outputs a 128-dimensional flow field microstructure feature vector after global average pooling. This vector is used to characterize the local shear rate and vortex intensity features of the tip leakage flow.

[0070] S2.2: Particle Interaction Branch (GNN Network Structure): The input layer receives data centered on the target particle with a radius R = 15d. p (d) pThe system constructs a local particle topology graph (containing node feature matrices and adjacency matrices) based on particle diameter, defining particles as nodes and inter-particle distances as edges. A two-layer graph attention layer (GAT) is used. The first layer employs a multi-head attention mechanism (8 attention heads) to extract high-order collision correlation features of neighboring nodes; the second layer uses a single-head attention mechanism for feature aggregation and dimensionality reduction. The output layer uses a Sum-Pooling operation to output a 64-dimensional particle interaction feature vector, which is used to quantify the aggregation degree of the particle swarm within the pump and the wake shielding effect.

[0071] S3: Cross-fuse the multi-source feature vectors obtained in S2 to obtain a multimodal fused feature vector, and establish the Euler-Lagrange bidirectional coupled partial differential control equations. Specifically:

[0072] S3.1: An adaptive attention fusion module is used to dynamically concatenate the flow field microstructure feature vector and the particle interaction feature vector to generate a multimodal fusion feature vector. Specifically, the adaptive attention module automatically calculates the weights of the contributions of the core vortex region and the high shear region of the flow field to particle collisions and sedimentation, thereby updating the weight allocation of the flow field microstructure feature vector and the particle interaction feature vector, and then dynamically concatenates them to generate a weighted multimodal fusion feature vector.

[0073] S3.2: Establish the partial differential governing equations for the Euler-Lagrange two-way coupling, including:

[0074] For the incompressible fluid phase inside the pump, the continuity equation and momentum conservation (NS) equation are derived as follows:

[0075]

[0076]

[0077] In the formula, It is a vector differential operator used to represent spatial differential operations such as gradient, divergence, or Laplace; For fluid velocity, For fluid pressure, For time, and These are fluid density and dynamic viscosity, respectively. This is the momentum source term for the particle group acting on the fluid.

[0078] For Lagrangian granular phases, traditional macroscopic static empirical formulas (such as the Hertz-Mindlin contact model) are abandoned, and Newton's second law is introduced to construct particle dynamics equations, as follows:

[0079]

[0080] In the formula, For particle mass, Instantaneous settlement velocity, For time, For alternating traction, Net weight and buoyancy, For virtual mass force; For particle collision contact force, in this embodiment of the invention, the particle interaction feature vector output by the particle interaction branch is denoted as h. p Map it to particle collision contact force ,Right now:

[0081]

[0082] In the formula, The network is a mapping network consisting of fully connected layers and nonlinear activation functions, with the following parameters: And introduce the local irreversible entropy yield ( The equation serves as a physical constraint on energy dissipation, with theoretical derivation replacing empirical fitting:

[0083]

[0084] In the formula, Let be the viscous dissipation function. The energy dissipation rate caused by inelastic particle collisions; T is the thermodynamic temperature, which is the local absolute temperature here.

[0085] S4: Input the multimodal fusion feature vector obtained in S3 into the physical information neural network model, and output the instantaneous particle settling velocity, particle collision contact force, local entropy yield, and globally reconstructed three-dimensional flow field. For example... Figure 2 As shown.

[0086] This step utilizes a deep learning model with embedded physical constraints to achieve accurate inversion of collision and settlement characteristics. The physical information neural network (PINN) backend inference structure and training strategy based on a gradient-aware adaptive physical loss function employed in this embodiment are as follows:

[0087] Input layer: Receives a 192-dimensional multimodal fusion feature vector (128-dimensional flow field microstructure features + 64-dimensional particle interaction features) concatenated by an adaptive attention module, and cascades the instantaneous Reynolds number Re of the particles. p With relative slip velocity u slip Both can be determined by the fluid velocity u. f and particle instantaneous velocity v p (Calculated from particle trajectory tracking results) That is:

[0088]

[0089]

[0090] Fully connected layers: 4 layers in total. Layer 1 has 256 neurons (Dropout rate 0.2), Layer 2 has 128 neurons (Dropout rate 0.2), Layer 3 has 64 neurons (Dropout rate 0.1), and Layer 4 has 32 neurons (Dropout rate 0). The activation functions used are all second-order continuously differentiable Swish functions to meet the smoothness requirements when calculating the Hessian matrix using automatic differentiation in PINN.

[0091] Output layer: Outputs the predicted instantaneous particle settling velocity, the entropy production rate (local entropy production rate) related to the particle collision contact force field and energy dissipation, and auxiliary variables (global reconstruction of the three-dimensional flow field) for calculating the physical residual.

[0092] Model Training and Physical Constraints: After each forward propagation, the automatic differentiation interface is invoked to construct the residuals of the fluid continuity equation, the Navier-Stokes equation, the particle dynamics equation, and the entropy yield equation as an energy constraint. The optimizer employs a hybrid optimization algorithm combining Adam and L-BFGS, constructing the following global-local collaborative total loss function, i.e., a gradient-aware adaptive physical loss function. The formula is:

[0093]

[0094] in, , , , , , All are coefficients. The data error between the model predictions and the particle analytical direct numerical simulation (PR-DNS); , These are the residuals of the fluid continuity equation and the Navier-Stokes equation, respectively. The residuals of the particle dynamics equations substituted into the model predictions; The residual of the entropy productivity equation; For the particle dynamics topological constraint residuals in the particle interaction branch; For a dynamic adaptive weight tensor, x represents the spatial coordinates of the flow field, and t represents time.

[0095] The adaptive physics loss function introduces the local flow field velocity gradient tensor. Positively correlated dynamic adaptive weight tensor During training iterations, the physical equation loss weights of the core region of the pump tip leakage vortex and the flow field stripping region (gradient maxima) are automatically amplified, forcing the physical information neural network to strictly adhere to the laws of conservation of mass, momentum, and energy in the strong shear region.

[0096] S5: Based on the output of S4, calculate the pump performance factor, particle settling risk, particle erosion wear risk, and flow field stability index, and perform graded adjustments accordingly, such as... Figure 3 As shown. Among them, the pump performance factor η is used to characterize the degree of deviation of the pump's current operating performance from its rated operating conditions, and the particle settling risk R... Settle R is used to characterize the likelihood of particle deposition and the risk of particle erosion wear. wear Used to characterize the intensity of erosion wear caused by particles on flow-through components, flow field stability S flow These indicators are used to characterize the degree of unsteady fluctuations in the flow field within the pump. In this embodiment, each indicator can be calculated as follows:

[0097]

[0098] Where H, Q, and P represent the current head, flow rate, and input power, respectively, and H0, Q0, and P0 represent the corresponding parameters under rated operating conditions; v s v is the particle settling velocity. c The local critical levitation velocity; F col E represents the contact force during particle collisions. dis F represents the local collision energy dissipation value. ref and E ref For reference operating conditions, the corresponding values ​​are as follows; , , The standard deviations of fluctuations in pressure, velocity, and entropy production rate, respectively, p ref u ref S ref For reference operating conditions; , , These are the weighting coefficients. The standard deviation measures the dispersion of a set of data.

[0099]

[0100] in, Let x be the population standard deviation, N be the total number of data points in the population, and x be the population standard deviation. i Each specific data point, The average value of the overall data.

[0101] The logic for the four-level indicator linkage assessment and adaptive adjustment is as follows:

[0102] Level 1 performance compensation (efficiency-oriented): When the pump performance factor When the value is below the threshold of 1 (set to 0.90 in this embodiment), it is determined that the hydraulic performance has deteriorated or the output power is insufficient. At this time, the frequency converter is driven to increase the speed to compensate for the loss of head or flow.

[0103] Secondary sedimentation intervention (anti-clogging guidance): When the risk of particle settling R Settle When the threshold value exceeds 2 (set to 1.00 in this embodiment), it is determined that the local flow field is insufficient to effectively carry particles. At this time, the guide vane angle is adjusted to improve the suspension state of particles by optimizing the internal flow channel structure and prevent accumulation and blockage.

[0104] Level 3 Wear Protection (Life-Oriented): When the risk of particle erosion wear is R... wear When the threshold value exceeds 3 (set to 1.20 in this embodiment), it is determined that the particle kinetic energy is too large, causing serious damage to the flow components. At this time, the rotation speed is adjusted (the rotation speed is reduced) to reduce the impact intensity of the particles on the wall and extend the equipment life.

[0105] Level 4 Instability Warning (Stability-Oriented): When the smooth stability S... flow When the threshold of 4 is exceeded (0.30 in this embodiment), it is determined that the system has entered an unsteady operating zone (such as surge or flow separation). At this time, the combined speed and guide vane adjustment program is triggered to forcibly rebuild a stable flow field and send an abnormal operation warning signal to the central control room.

[0106] A system for predicting particle settling, collision, and flow characteristics within a pump includes:

[0107] A multi-source heterogeneous data acquisition module is used to acquire real-time video streams of high-speed flow fields and particle motion trajectory signals inside the pump, and integrates a high-fidelity numerical simulation data interface.

[0108] The multi-scale feature extraction module is used to extract flow field microstructure features (including Q-criterion) and particle topological interaction features in parallel using the flow field perception branch and particle interaction branch, respectively, to generate a multi-modal fusion feature vector;

[0109] The physical information fusion prediction module is used to input the feature vector into a pre-trained physical information neural network based on gradient perception and adaptive physical loss function, and predict instantaneous collision force and settling velocity, local entropy yield and global reconstruction of three-dimensional flow field under the physical constraints of the Navier-Stokes equations and dynamic laws.

[0110] The operational condition quantitative assessment module is used to construct the pump performance factor η and particle settling risk R based on the instantaneous particle settling velocity, particle collision contact force, local entropy yield, and globally reconstructed three-dimensional flow field output by the physical information fusion prediction module. Settle Risk of particle erosion and wear (R) wear and flow field stability Sflow This allows for a quantitative assessment of the flow state within the pump, particle deposition trends, erosion and wear levels, and the risk of unsteady operation, thus quantifying the flow state and wear risk.

[0111] The intelligent feedback control module is used to execute graded alarms based on the evaluation results and to adjust equipment operating parameters such as pump speed and guide vane angle in conjunction with the results.

[0112] Example 2

[0113] The embodiment uses the tip leakage flow region of a multi-stage gas-liquid mixed-transfer pump as a typical application scenario. The specific implementation steps of the method for predicting particle settling and collision and flow characteristics within the pump are as follows:

[0114] S01: Construct a training dataset for real-time acquisition environment and virtual-real fusion multiphase flow. This step aims to establish a high-fidelity data foundation that maps physical experiments to virtual simulations.

[0115] In this embodiment, a non-invasive optical acquisition unit is first deployed on the side wall of the transparent multi-stage mixed-transport pump test section. This unit consists of two Phantom TMX series high-resolution high-speed industrial cameras arranged at a 90-degree angle and a matching macro lens. The sampling frequency is set to 2000 frames / second and the resolution is 2048×2048 pixels. With the help of a 10W high-frequency continuous wave laser sheet light source extended by a cylindrical lens, a precise tomographic scan of the highly turbulent shear region of the pump impeller leading edge and blade tip gap is performed. The camera and laser are synchronously triggered at the microsecond level through TTL signals, and the massive image data is transmitted in real time to an edge computing workstation equipped with an NVIDIA A100 GPU through the CoaXPress 2.0 interface.

[0116] Meanwhile, a virtual data domain was constructed using particle analytical direct numerical simulation (PR-DNS) technology to solve the Navier-Stokes equations at a grid resolution of tens of millions, simulating Reynolds numbers in the range of 10. 4 Up to 10 6 The turbulent flow field of the tip leakage vortex between the particles was analyzed, and the particle settling velocity and particle collision contact force were obtained by integrating over the particle surface. Furthermore, based on the change in kinetic energy before and after the collision and / or the energy loss during the contact process, the collision energy dissipation value of the particles in the tip leakage vortex region was calculated. Subsequently, the above flow field contour map, particle settling velocity, particle collision contact force, and collision energy dissipation value were spatiotemporally registered to generate a high-fidelity pre-training sample library, providing supervised labels and physical constraint references with physical truth values ​​for the subsequent physical neural network.

[0117] S02: Preprocessing and feature extraction are performed on the image data, particle trajectory data, and simulation label data in the training sample library. Specifically, image data undergoes denoising, calibration, cropping, and normalization; particle trajectory data undergoes outlier removal, trajectory association, interpolation completion, motion parameter calculation, and time window slicing; simulation label data undergoes time alignment, spatial registration, and dimensionless processing; and a graph structure data is constructed for the target particle neighborhood based on the particle trajectory data. After preprocessing, flow field tensor features and particle topological features are obtained.

[0118] This embodiment employs a two-stream deep neural network architecture for parallel feature extraction, with the specific structure as follows:

[0119] (1) Flow field perception branch (CNN network structure):

[0120] Input layer: Receives the preprocessed local flow field slice tensor centered on the target particle (including the TKE channel and the Q-criterion eddy feature channel pre-computed by the physical feature constructor).

[0121] Convolutional layers: Based on a custom ResNet-18 variant architecture. The first layer is a convolutional kernel (stride 2, padding 3), followed by 4 residual blocks. Each residual block contains 2 convolutional layers and a batch normalization layer, with the number of channels increasing to 32, 64, and 128 layer by layer.

[0122] Output layer: The global average pooling layer is followed by the Flatten layer, which outputs a 128-dimensional flow field microstructure feature vector. This vector implicitly contains the local shear rate and vortex intensity characteristics of the tip leakage flow.

[0123] (2) Particle-interaction branch (GNN network structure):

[0124] Input layer: Receives the constructed local particle topology graph data, including node feature matrices (containing relative positions and velocities) and adjacency matrices (indicated by...). (Constructed for radius)

[0125] Graph Attention Layer (GAT): 2 layers in total. The first layer uses a multi-head attention mechanism (8 attention heads) to extract collision correlation features of neighbor nodes in a high-gradient flow field; the second layer uses a single-head attention mechanism for feature aggregation and dimensionality reduction.

[0126] Output layer: The Sum-Pooling operation is used to output a 64-dimensional particle interaction feature vector, which quantifies the degree of particle aggregation and wake shielding effect in the pump.

[0127] S03: Physical information fusion and parameter prediction.

[0128] This step utilizes a deep learning model with embedded physical constraints to achieve accurate inversion of collision and settlement characteristics. The Physical Information Neural Network (PINN) backend inference structure and training strategy used in this embodiment are as follows:

[0129] Input layer: Receives a 192-dimensional multimodal fusion feature vector (128-dimensional flow field microstructure features + 64-dimensional particle interaction features) concatenated by an adaptive attention module, and cascades the instantaneous Reynolds number Re of the particles. p With relative slip velocity u slip ;

[0130] Fully connected layers: 4 layers in total. Layer 1 has 256 neurons (Dropout rate 0.2), Layer 2 has 128 neurons (Dropout rate 0.2), Layer 3 has 64 neurons (Dropout rate 0.1), and Layer 4 has 32 neurons (Dropout rate 0). The activation functions used are all second-order continuously differentiable Swish functions to meet the smoothness requirements when calculating the Hessian matrix using automatic differentiation in the PINN network.

[0131] Output layer: Outputs the predicted instantaneous particle settling velocity, the entropy production rate associated with the particle collision contact force field and energy dissipation, and auxiliary variables used to calculate the physical residual (global reconstruction of the three-dimensional flow field).

[0132] Model training and physical constraints: After each forward propagation, the system calls the automatic differentiation interface to calculate the partial derivatives of the flow field variables output by the network, and constructs the fluid continuity equation residuals. Navier-Stokes equation residuals Particle dynamics equation residuals and the residual of the entropy production rate equation Specifically, a dynamic adaptive weight tensor that is positively correlated with the local flow field velocity gradient is introduced. Construct the total loss function:

[0133]

[0134] The optimizer is Adam, and the initial learning rate is set to 0.001. By minimizing the physical residuals, the network output is forced to satisfy the laws of hydrodynamic conservation, thereby ensuring that the model can still output predicted values ​​that conform to physical laws under complex conditions where experimental ground truth labels are lacking.

[0135] S04: Quantitative Assessment of Operating Conditions and Characteristic Evolution Analysis. This step aims to transform the microscopic dynamic parameters obtained from the inversion into quantifiable process control indicators, providing data support for closed-loop regulation.

[0136] This step aims to transform the instantaneous particle settling velocity, particle collision contact force, local entropy yield, and globally reconstructed three-dimensional flow field output by SO3 into quantifiable process control indicators, providing data support for closed-loop regulation. Through multi-parameter fusion, four core evaluation indicators are ultimately generated: pump performance factor. Particle sedimentation risk R Settle Risk of particle erosion and wear (R) wear and flow field stability S flow This is used to quantitatively reveal the evolution of pump characteristics.

[0137] S05: Intelligent Hierarchical Early Warning and Closed-Loop Control. This step executes an adaptive closed-loop adjustment strategy based on the four indicators output from S04, achieving a dynamic balance between equipment performance, lifespan, and stability through threshold judgment logic. The specific adjustment logic is as follows:

[0138] Performance compensation logic (Level 1): When the pump performance factor is detected... When the output is below the preset threshold of 1 (set to 0.90 in this embodiment), it is determined that the hydraulic output is insufficient. The operating speed is increased by the frequency converter through the industrial bus command to maintain the head and flow rate required by the process.

[0139] Deposition Intervention Logic (Level 2): ​​When a local particle settling risk R is detected Settle When the threshold value exceeds 2 (set to 1.00 in this embodiment), it is determined that there is a potential dead zone deposition risk, and the guide vane angle is automatically adjusted to enhance the particle carrying capacity by changing the internal flow field structure.

[0140] Wear Protection Logic (Level 3): When particle erosion wear risk R is detected wear When the threshold value exceeds 3 (set to 1.20 in this embodiment), it is determined that there is a serious risk of erosion in the impeller or flow components, and the frequency converter is instructed to perform a speed reduction operation to reduce the kinetic energy of particle collision and extend the hardware life.

[0141] Instability warning logic (level 4): When the flow field stability S is detected... flow When the threshold of 4 is exceeded (set to 0.3 in this embodiment), a stall or agglomeration anomaly is determined to have occurred. The combined speed and guide vane adjustment program is triggered to reconstruct the flow regime, and a red warning signal is simultaneously sent to the central control room.

[0142] To verify the advantages of the method for predicting particle settling, collision, and flow characteristics in the pump compared to conventional techniques, a comparative experiment was conducted in the tip leakage flow region of a multi-stage gas-liquid mixed-transfer pump.

[0143] Experimental conditions and evaluation indicators: Pump speed was set at 1450 rpm, inlet solids content was set at 5%, and median particle diameter d was set at 5%. 50= 0.5 mm. The results of the particle-resolved direct numerical simulation (PR-DNS) method based on a fully implicit two-way coupling scheme are used as the baseline true values. Evaluation index selection: root mean square error of particle settling velocity (RMSE) v ), the relative error of the maximum collision contact force (E) Fmax The proportion of non-physical predictions and the time consumed in a single prediction step were compared with the embodiment of this invention (CNN+GNN+PINN physical deep cross-coupling model). The results are shown in Table 1. Three control groups were set up: traditional CFD-DEM empirical formula method (Hertz-Mindlin contact model + Wen-Yu drag model), pure data-driven CNN+GNN model (without embedded PINN physical equation constraints), and conventional pure PINN model (without CNN flow field feature pre-construction and GNN topology extraction).

[0144] Table 1 Comparative Experimental Results

[0145]

[0146] Traditional CFD-DEM empirical formula methods rely on static empirical formulas, making them ill-suited for unsteady flow fields with strong vortices, resulting in a maximum relative error of 35.8% in the maximum collision contact force. Pure data-driven CNN+GNN models, lacking physical constraints, exhibit severe non-physical prediction phenomena (as shown in Table 1, where the prediction results on the test set were tested by substituting them into the continuity and momentum equations, revealing that up to 14.5% of the predictions violated the laws of conservation of mass and momentum). This invention's method, by introducing a Physics-Informed Loss function, fundamentally eliminates these non-physical prediction phenomena, reducing the violation rate of conservation laws to 0%. Furthermore, conventional pure PINN models, lacking multi-scale feature pre-extraction of the local flow field, suffer from significantly limited prediction accuracy in complex flow field scenarios such as tip leakage flow; while this invention's method employs a multi-scale feature fusion dual-branch network, drastically compressing the root mean square error of particle settling velocity to 0.05 m / s, improving the collision force prediction accuracy by nearly 7 times compared to the conventional pure PINN model. Furthermore, while maintaining strict conservation of physical mechanisms and high fidelity, the method of this invention reduces the single-step prediction time from 14,400 seconds in traditional methods to 0.12 seconds, achieving an exponential leap in computational efficiency and effectively meeting the real-time requirements of online quantitative assessment and intelligent early warning in industrial settings.

[0147] The embodiments described above are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for predicting particle settling, collision, and flow characteristics within a pump, characterized in that, Includes the following steps: Step S1: Construct a training sample library for multiphase flow that integrates real-time acquisition environment with virtual-real fusion, and acquire flow field image data, particle trajectory data and simulation label data; Step S2: Preprocess the flow field image data, particle trajectory data and simulation label data obtained in Step S1 to obtain flow field tensor features and particle topological features. Use a dual-flow deep neural network architecture to extract the features in parallel to obtain multi-source feature vectors. The multi-source feature vectors include flow field microstructure feature vectors and particle interaction feature vectors. Step S3: Cross-fuse the multi-source feature vectors obtained in step S2 to obtain multi-modal fused feature vectors, and establish the Euler-Lagrange bidirectional coupled partial differential control equations; Step S4: Input the multimodal fusion feature vector obtained in step S3 into the physical information neural network model based on gradient-aware adaptive physical loss function, and output the instantaneous particle settling velocity, particle collision contact force, local entropy yield and global reconstructed three-dimensional flow field. Step S5: Calculate the pump performance factor, particle settling risk, particle erosion wear risk, and flow field stability index based on the output results of step S4, and perform graded adjustment accordingly.

2. The method for predicting particle settling, collision, and flow characteristics within a pump according to claim 1, characterized in that, The dual-flow deep neural network architecture includes a flow field perception branch and a particle interaction branch.

3. The method for predicting particle settling, collision, and flow characteristics within a pump according to claim 2, characterized in that, The flow field sensing branch structure is as follows: Input layer: Receives preprocessed local flow field slice tensors centered on the target particle; Convolutional layers: Based on a custom ResNet variant structure, it contains 4 residual modules. Each residual module contains 2 convolutional layers and a batch normalization layer. The first two residual modules have 32 output channels, the third residual module has 64 output channels, and the fourth residual module has 128 output channels. Output layer: After global average pooling, a 128-dimensional flow field microstructure feature vector is output, which is used to characterize the local shear rate and vortex intensity of the tip leakage flow.

4. The method for predicting particle settling, collision, and flow characteristics within a pump according to claim 3, characterized in that, The particle interaction branch structure is as follows: Input layer: Receives data centered on the target particle with a radius R = 15d. p The constructed local particle topology map data, where d p The diameter of the particle; Graph Attention Layer: There are 2 layers in total. The first layer uses a multi-head attention mechanism to extract high-order collision association features of neighbor nodes; the second layer uses a single-head attention mechanism for feature aggregation and dimensionality reduction. Output layer: The Sum-Pooling operation is used to output a 64-dimensional particle interaction feature vector, which is used to quantify the degree of particle aggregation and wake shielding effect in the pump.

5. The method for predicting particle settling, collision, and flow characteristics within a pump according to claim 4, characterized in that, The local flow field slice tensor incorporates the spatial distribution of the flow field Q criterion and the turbulent kinetic energy channels calculated by the physical feature constructor; the local particle topology graph data defines particles as nodes and the distance between particles as edges.

6. The method for predicting particle settling, collision, and flow characteristics within a pump according to claim 1, characterized in that, The process of generating the multimodal fusion feature vector is as follows: the adaptive attention module automatically calculates the weights of the core vortex region and the high shear region of the flow field on the contribution of particle collision and sedimentation, thereby updating the weight allocation of the flow field microstructure feature vector and the particle interaction feature vector, and then dynamically splicing to generate the weighted multimodal fusion feature vector.

7. The method for predicting particle settling, collision, and flow characteristics within a pump according to claim 1, characterized in that, The Euler-Lagrange bidirectional coupled partial differential governing equations include: The continuity equation and the Navier-Stokes equation are as follows: In the formula, It is a vector differential operator; For fluid velocity, For fluid pressure, For time, and These are fluid density and dynamic viscosity, respectively. This is the momentum source term for the particle group acting on the fluid; The particle dynamics equation is as follows: In the formula, For particle mass, Instantaneous settlement velocity, For alternating traction, Net weight and buoyancy, For virtual mass force; This refers to the contact force during particle collisions. Particle collision contact force The particle interaction feature vector output by the particle interaction branch is denoted as h. p The mapping yields, i.e.: In the formula, The network is a mapping network consisting of fully connected layers and nonlinear activation functions, with the following parameters: ; and local irreversible entropy production The equation, serving as the physical constraint for energy dissipation, is expressed as follows: In the formula, Let be the viscous dissipation function. The energy dissipation rate caused by inelastic particle collisions; T is the thermodynamic temperature.

8. The method for predicting particle settling, collision, and flow characteristics within a pump according to claim 7, characterized in that, The specific structure of the physical information neural network model is as follows: Input layer: Receives multimodal fused feature vectors and cascades granular instantaneous Reynolds numbers Re. p With relative slip velocity u slip ; Fully connected layers: 4 layers in total. Layer 1 has 256 neurons with a dropout rate of 0.2; Layer 2 has 128 neurons with a dropout rate of 0.2; Layer 3 has 64 neurons with a dropout rate of 0.1; Layer 4 has 32 neurons with a dropout rate of 0. The activation function used for all layers is the second-order continuously differentiable Swish function. Output layer: Outputs predicted instantaneous particle settling velocity, particle collision contact force field, local entropy yield, and globally reconstructed 3D flow field; Model Training and Physical Constraints: After each forward propagation, the automatic differentiation interface is called to construct the residuals of the fluid continuity equation, the Navier-Stokes equation, the particle dynamics equation, and the entropy yield equation as an energy constraint. The optimizer employs a hybrid optimization algorithm combining Adam and L-BFGS to construct a gradient-aware adaptive physical loss function. The formula is: in, , , , , , All are coefficients. The data error between the model predictions and the direct numerical simulation of particle analysis; , These are the residuals of the fluid continuity equation and the Navier-Stokes equation, respectively. The residuals of the particle dynamics equations substituted into the model predictions; The residual of the entropy productivity equation; For the particle dynamics topological constraint residuals in the particle interaction branch; For a dynamic adaptive weight tensor, x represents the spatial coordinates of the flow field, and t represents time.

9. The method for predicting particle settling, collision, and flow characteristics within a pump according to claim 1, characterized in that, The specific logic of the hierarchical adjustment is as follows: Level 1 performance compensation: When the pump performance factor When the value is below 0.90, it is determined that the hydraulic performance has deteriorated or the output power is insufficient. At this time, the frequency converter is driven to increase the speed to compensate for the head or flow loss. Secondary sedimentation intervention: when the particle settling risk R Settle When the value exceeds 1.00, it is determined that the local flow field is insufficient to effectively carry particles. At this time, the guide vane angle is adjusted to improve the suspension state of particles by optimizing the internal flow channel structure and prevent accumulation and blockage. Level 3 wear protection: When the risk of particle erosion wear is R wear When the value exceeds 1.20, it is determined that the particle kinetic energy is too large, causing serious damage to the flow components. At this time, the rotation speed should be adjusted to reduce the impact intensity of the particles on the wall and extend the equipment life. Level 4 Instability Warning: When the smooth stability S flow When the value exceeds 0.30, it is determined that the system has entered the unsteady operation zone. At this time, the combined speed and guide vane adjustment program is triggered to forcibly rebuild the stable flow field and send an abnormal operation warning signal to the central control room.

10. A system for predicting particle settling, collision, and flow characteristics in a pump, used to execute the method for predicting particle settling, collision, and flow characteristics in a pump as described in claims 1-9, characterized in that, include: A multi-source heterogeneous data acquisition module is used to acquire real-time video streams of high-speed flow fields and particle motion trajectory signals inside the pump, and integrates a high-fidelity numerical simulation data interface. The multi-scale feature extraction module is used to extract the microstructure features of the flow field and the topological interaction features of the particles in parallel using the flow field perception branch and the particle interaction branch, respectively, to generate a multimodal fusion feature vector. The physical information fusion prediction module is used to input the feature vector into a pre-trained physical information neural network based on gradient perception and adaptive physical loss function, and predict instantaneous collision force and settling velocity, local entropy yield and global reconstruction of three-dimensional flow field under the physical constraints of the Navier-Stokes equations and dynamic laws. The operational condition quantitative assessment module is used to construct the pump performance factor η and particle settling risk R based on the instantaneous particle settling velocity, particle collision contact force, local entropy yield, and globally reconstructed three-dimensional flow field output by the physical information fusion prediction module. Settle Risk of particle erosion and wear (R) wear and flow field stability S flow This allows for a quantitative assessment of the flow state within the pump, particle deposition trends, erosion and wear levels, and the risk of unsteady operation, thus quantifying the flow state and wear risk. The intelligent feedback control module is used to execute graded alarms based on the evaluation results and to adjust equipment operating parameters such as pump speed and guide vane angle in conjunction with the results.

Images