Method for rapidly predicting particle rotation of hydrocyclone based on data driving

Abstract

The invention relates to a method for rapidly predicting particle rotation of a cyclone based on data driving, which comprises the following steps of: carrying out CFD numerical simulation through an Euler-Lagrange coupling model, combining PIV (particle image velocimetry) measurement and high-speed camera experiment data, and dynamically fusing multi-source data by adopting Kalman filtering to construct a mixed data set; in the feature engineering stage, thousand-dimensional flow field data are compressed to 50-dimensional through PCA dimension reduction and an auto-encoder, and key physical quantities such as centrifugal acceleration, shear stress and vorticity are weighted and fused. A multi-model collaborative prediction system (BP / XGBoost / CATBoost / RF / AdaBoost / SVM) is innovatively constructed, and algorithm advantage complementation is realized through dynamic weight fusion (error reciprocal distribution + abnormal weight drop). Classified hyper-parameter optimization (Bayesian optimization tree model depth / learning rate, grid search SVM kernel parameters) is adopted, TensorRT quantization (FP16) and ONNX conversion are combined, and the inference speed of embedded deployment reaches 48 ms. And the prediction result drives the PID controller to accurately adjust the inlet flow in real time. And through data-driven modeling and interpretability analysis, an efficient solution is provided for cyclone optimization control.

Description

Technical Field

[0001] The present invention relates to the technical field of fluid mechanics, and specifically to a fast prediction method for the self-rotation of cyclone particles based on data-driven. Background Art

[0002] Existing Technical Bottlenecks

[0003] Traditional CFD simulation: It relies on solving the Navier-Stokes equations for the flow field, requires high-resolution grid division, and takes a long time to calculate (usually from several hours to several days), making it difficult to meet the industrial real-time requirements.

[0004] Complexity of particle dynamics: In a cyclone, particles are affected by the coupling of multiple physical fields such as centrifugal force, drag force, and turbulent diffusion. The self-rotation speed has a strong non-linear relationship with parameters such as particle size, density, and vorticity of the flow field. Traditional simulation methods are difficult to accurately model.

[0005] Limitations of experimental measurements: High-speed photography or PIV technology is costly and difficult to achieve full flow field coverage in industrial-grade equipment.

[0006] Therefore, a fast prediction method for the self-rotation of cyclone particles based on data-driven is needed. Summary of the Invention

[0007] The present invention provides a fast prediction method for the self-rotation of cyclone particles based on data-driven to solve the problems of the existing technology.

[0008] To solve the above technical problems, the present invention is realized through the following technical solutions: A fast prediction method for the self-rotation of cyclone particles based on data-driven, including the following steps:

[0009] S1: Data acquisition: Through a computational fluid dynamics (CFD) simulation tool, numerically simulate the motion state of particles in the cyclone to obtain multi-dimensional data including flow field velocity distribution, particle trajectories, pressure gradients, and turbulent kinetic energy;

[0010] S1.1 Model selection and solution framework adopt the coupling framework of the Eulerian multiphase flow model and the discrete element method (DEM):

[0011] The fluid phase uses the Reynolds stress model (RSM) to solve anisotropic turbulence, and the transport equation includes 7 key parameters such as turbulent diffusion, molecular diffusion, and shear stress generation;

[0012] The particle phase solves translational and rotational motions through Newton's second law, and the Hertz-Mindlin (NoSlip) model is used for contact force calculation. The parameter settings are as follows: the particle-particle collision restitution coefficient is 0.35, the static friction coefficient is 0.55, and the rolling friction coefficient is 0.08; the particle-wall collision restitution coefficient is 0.5, the static friction coefficient is 0.3, and the rolling friction coefficient is 0.05;

[0013] The mesh generation is based on ICEM software to generate unstructured meshes:

[0014] The total number of meshes is 86,678 (verified by mesh independence, with the maximum tangential velocity deviation < 5%);

[0015] The encrypted area includes the wall boundary layer, tangential inlet, and column-cone junction area;

[0016] The mesh size strictly meets the requirements of CFD-DEM coupling (mesh size > particle diameter); Particle parameters

[0017] Material properties: POM plastic particles, density 1200 kg / m 3 , and three groups with particle sizes of 1 / 3 / 5 mm;

[0018] Injection method: Uniformly generated through the EDEM particle factory at the inlet section;

[0019] Surface treatment: Bicolor tracer particles (black and white spraying) are used for motion feature recognition;

[0020] Two-phase flow setting and solution strategy:

[0021] Fluent solves the steady-state flow field (time step 5×10 -4 s);

[0022] EDEM solves the particle motion (time step 2×10 -5 s);

[0023] The gas-solid two-way momentum exchange boundary condition is achieved through the coupling interface:

[0024] Inlet: Velocity-inlet (velocity range corresponding to 40 - 70 m 3 / h gas volume);

[0025] Outlet: Outflow (overflow / underflow split ratio 0 - 100%);

[0026] Pressure-velocity coupling: SIMPLE algorithm;

[0027] Turbulence intensity: I = 0.16(ReD)^(-1 / 8);

[0028] Dynamic calibration of the coupling iteration process:

[0029] After the Fluent flow field reaches macroscopic steady state, it is transferred to EDEM;

[0030] The particle motion information is fed back to the fluid phase through the momentum source term;

[0031] The total calculation duration is 4 s (meeting the requirements of gas-solid two-phase dynamic equilibrium).

[0032] S2: Feature Extraction and Preprocessing: Perform feature engineering on the data in step S1, including but not limited to dimensional normalization, noise filtering, and extraction of key physical quantities, and construct a standardized data set, where the key physical quantities include but not limited to centrifugal acceleration and local shear stress;

[0033] S2.1 Data Dimensionality Reduction

[0034] PCA Dimensionality Reduction: Retain the principal components with a cumulative variance contribution rate > 95%, and compress the original flow field data (such as a 1000-dimensional velocity vector field) to 50 dimensions;

[0035] Autoencoder Design: A 3-layer encoder (input - 512 - 256 - 50), ReLU activation, a symmetric decoder structure, and the loss function is the reconstruction mean square error.

[0036] S2.2 Physical Quantity Weighted Fusion

[0037] Calculation of Key Features:

[0038] Centrifugal acceleration;

[0039] Local shear stress;

[0040] Vorticity.

[0041] Weighting Strategy: Based on SHAP value analysis within a sliding time window (window length = 0.1T, T is the cyclone characteristic time scale), dynamically assign weights: centrifugal acceleration 0.3 - 0.5, shear stress 0.2 - 0.4, vorticity 0.2 - 0.4, and the weight update frequency is synchronized with the sensor sampling rate (10 kHz).

[0042] S2.3 Data Standardization

[0043] Noise Filtering: Remove high-frequency noise using wavelet transform (Daubechies-4 basis function);

[0044] Normalization: Min-Max normalize to [0, 1], and use Box-Cox transform for skewed distribution data (such as pressure gradient).

[0045] S3: Machine Learning Model Construction: Based on six types of models, namely integrated BP, CATBoost, RF, AdaBoost, XGBoost, and SVM, design a prediction model architecture and improve the model generalization ability through hyperparameter optimization;

[0046] S3.1 Integrated Model Architecture Design

[0047] Based on six types of models, namely BP (Backpropagation Neural Network), CATBoost (an efficient implementation of Gradient Boosting Tree Algorithm), RF (Random Forest), AdaBoost (Adaptive Boosting Algorithm), XGBoost (an efficient Gradient Boosting library), and SVM (Support Vector Machine), the design of the integrated prediction model architecture is carried out.

[0048] The specific integration method is as follows:

[0049] For the input standardized data set after feature extraction and preprocessing, it is respectively input into the above six types of models.

[0050] BP model: Construct a neural network with a suitable hidden layer structure (for example, the number of neurons in the hidden layer is adjusted according to data characteristics and model complexity). The input layer receives feature data, the hidden layer uses a suitable activation function (such as ReLU) for calculation, and the output layer outputs the values related to the predicted particle rotation speed.

[0051] CATBoost model: Set suitable parameters, such as the depth of the tree, learning rate, number of leaf nodes, etc. Input the feature data into the model and train it through the way of gradient boosting to obtain the prediction result.

[0052] RF model: Determine the number of decision trees according to the feature dimension and data scale. Randomly sample the input data and input it into each decision tree for training and prediction. Finally, integrate the prediction results of all decision trees (such as voting method or averaging method) to obtain the final predicted value.

[0053] AdaBoost model: Initialize the sample weights. According to the process of the AdaBoost algorithm, train multiple weak classifiers (such as decision stumps) in turn, adjust the sample weights according to the performance of each weak classifier, and finally combine all weak classifiers into a strong classifier to output the prediction result.

[0054] XGBoost model: Set the relevant parameters of the tree (such as maximum depth, minimum leaf node weight, etc.) and regularization parameters. Input the feature data into the model and accelerate the training process through parallel computing to obtain the predicted value of the particle rotation speed.

[0055] SVM model: Select a suitable kernel function (such as Radial Basis Function RBF, Polynomial Kernel Function, etc.). According to the data distribution and feature dimension, adjust the kernel function parameters and penalty parameter C. Input the feature data into the SVM model for training and prediction.

[0056] S3.2 Hyperparameter Optimization

[0057] Search space:

[0058] Learning rate: Log range [1e-4, 1e-2];

[0059] Number of LSTM units: 32 / 64 / 128;

[0060] Batch size: 32 / 64 / 128.

[0061] Bayesian optimization: Gaussian process surrogate model, iterate 50 times, select the parameter combination with the largest expected improvement (EI).

[0062] S4: Model training and optimization: Adopt a batch training strategy, dynamically adjust model parameters in combination with cross-validation technology, and evaluate model performance through a loss function until convergence, where the loss function uses mean squared error;

[0063] S4: Model training and optimization

[0064] S4.1 Batch training

[0065] Data division: 70% training set, 15% validation set, 15% test set;

[0066] Training configuration: Adam optimizer, initial learning rate 0.001, cosine annealing schedule, batch size 64.

[0067] S4.2 Overfitting suppression

[0068] Early stopping mechanism: Monitor the loss of the validation set, and terminate training if it does not decrease for 10 consecutive epochs;

[0069] Regularization: L2 regularization (λ = 0.001), Dropout rate 0.2 (fully connected layer)

[0070] S5: Practical application verification: Deploy the trained model to an industrial hydrocyclone device, input sensor or simulated data in real time, output the prediction result of the particle rotation speed, and conduct an error comparison analysis with the measured data;

[0071] S5.1 Embedded deployment

[0072] Model lightweighting: Quantize to FP16 with TensorRT, and compress the model volume to 30% of the original size;

[0073] Hardware configuration: Jetson Xavier NX, inference latency <50ms.

[0074] S5.2 Control linkage

[0075] Real-time prediction: Receive sensor data (pressure, flowmeter) through the Modbus protocol, and output the prediction value every 200ms;

[0076] Parameter adjustment: If the predicted rotation speed deviates from the set value by ±15%, the PID controller adjusts the inlet flow rate of the cyclone (adjustment range ΔQ = K_p × e(t)).

[0077] S5.3 Error analysis

[0078] Indicators: Mean Absolute Error (MAE) < 5 rad / s, coefficient of determination R 2 > 0.92;

[0079] Failure handling: When the prediction error exceeds the threshold for 5 consecutive times, switch to the CFD online simulation mode and trigger an alarm.

[0080] In a specific embodiment, the CFD simulation tool in step S1 adopts the Euler-Lagrange coupling model, and the specific implementation includes:

[0081] The particle phase is modeled using the Discrete Element Method (DEM), the particle size range is 1 - 5 mm, and the density is 1200 kg / m 3 ;

[0082] The fluid phase uses the Reynolds Stress Model (RSM) to describe anisotropic turbulence;

[0083] The gas-solid two-phase coupling adopts the Hertz-Mindlin (NoSlip) contact model;

[0084] Set the particle-wall collision restitution coefficient to 0.5 and the static friction coefficient to 0.3;

[0085] The inlet boundary condition is a velocity inlet, and the outlet uses a free outflow boundary;

[0086] The time step is set to the coupling ratio of Fluent 5×10 -4 s and EDEM 2×10 -5 s.

[0087] In a specific embodiment, the feature extraction in step S2 includes:

[0088] Dimensionality reduction of high-dimensional data is performed by Principal Component Analysis (PCA) or autoencoders;

[0089] The operating parameters and structural parameters affecting the rotation speed of the particles are weighted and fused.

[0090] In a specific embodiment, in the model construction of step S3, a hybrid architecture of six machine learning models including BP neural network, CATBoost, RF, AdaBoost, XGBoost, and SVM is adopted.

[0091] In a specific embodiment, during the optimization process of step S4, the Bayesian optimization algorithm is used to globally search for hyperparameters, and an early stopping mechanism (Early Stopping) is introduced to prevent overfitting.

[0092] In a specific embodiment, in the practical application of step S5, the model is lightweight deployed through an embedded system and linked with the cyclone control unit to adjust the separation efficiency parameters in real time.

[0093] In a specific embodiment, in step S1, the CFD simulation data and the experimental data are jointly calibrated through a dynamic weight fusion strategy, specifically including:

[0094] Verify the time-space synchronization of the experimental data (PIV measurement, high-speed camera trajectory tracking), and eliminate abnormal sampling points;

[0095] Use the Kalman filter algorithm to fuse the flow field velocity distribution simulated by CFD and the experimental measurement values, and dynamically adjust the confidence weights;

[0096] When constructing the mixed dataset, give priority to using CFD data for gas-liquid two-phase flow conditions and experimental data for liquid-solid two-phase flow conditions.

[0097] The beneficial effects of the present invention are as follows:

[0098] 1. The present invention adopts dynamic fusion of CFD simulation and experimental data (Kalman filter weight assignment), combined with model lightweighting (quantizing TensorRT to FP16), to achieve an inference speed of 48 ms for embedded deployment, which improves the CFD calculation efficiency by 2-3 orders of magnitude compared with the traditional method, meeting the requirements of industrial on-line control. For liquid-solid two-phase flow conditions, give priority to using experimental data for calibration, and the experimental weight in the high-turbulence area is increased to 0.7 to ensure the model stability under extreme conditions;

[0099] 2. Construct a six-model hybrid architecture (BP / XGBoost / CATBoost / RF / AdaBoost / SVM), and achieve complementary advantages of algorithms through dynamic weight fusion (error reciprocal assignment + abnormal weight reduction). The determination coefficient R of the integrated model 2 reaches 0.95, which is more than 6% higher than that of a single model. Feature engineering reduces the dimension through PCA + autoencoder (retaining 95% variance), and combines SHAP values to dynamically weight and fuse key physical quantities such as centrifugal acceleration (0.3-0.5) and shear stress (0.2-0.4), so that the prediction error is stably controlled within 4 rad / s, and the error of a single CFD simulation is reduced by 60%;

[0100] 3. The predicted results are used to drive the PID controller to adjust the inlet flow rate in real time (ΔQ = K_p × e(t)), achieving precise control of the particle self-rotation speed deviation within ±15%. In industrial applications, the separation efficiency is increased by 15%, and the energy consumption is reduced by 8 - 12%. The model volume is compressed to 30% of the original size, and the single-node deployment cost is reduced by 70% based on the Jetson Xavier NX hardware. It supports parallel control of multiple cyclone clusters, combining technological advancement and economic feasibility. Description of the Drawings

[0101] Figure 1 It is a schematic structural diagram of the cyclone for CFD data simulation of the present invention.

[0102] Figure 2 It is a schematic diagram of the overall process of the present invention.

[0103] Figure 3 It is a schematic diagram of the multivariate correlation analysis of density (ρ) and feature importance (SHAP value) in the cyclone particle self-rotation prediction model of the present invention.

[0104] Figure 4 It is a schematic diagram for analyzing the influence of the contribution degree of multi-parameter SHAP values in the cyclone particle self-rotation prediction model of the present invention.

[0105] Figure 5 It is a schematic diagram for analyzing the global influence heat map of multiple features of the cyclone particle self-rotation prediction model based on SHAP values of the present invention.

[0106] Figure 6 It is a schematic diagram for ranking the average influence degree of the SHAP values of the key parameters in the cyclone particle self-rotation prediction model of the present invention.

[0107] Figure 7 It is a schematic diagram for analyzing the scatter of the influence direction and intensity distribution of multiple features of the cyclone particle self-rotation prediction model based on SHAP values of the present invention.

[0108] Figure 8 It is a schematic diagram for decomposing the positive and negative contributions of features of the cyclone particle self-rotation prediction model based on SHAP values of the present invention. Detailed Embodiments

[0109] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0110] Such as Figures 1 to 8 The data-driven rapid prediction method for cyclone particle self-rotation shown.

[0111] As shown in Figure 3 Figure [FIGURE NUMBER]: The contribution degree distribution of particle density (ρ) to the rotational speed prediction model is shown. The SHAP value shows a piecewise linear upward trend with the increase of density (the slope is significant when 600 ≤ ρ ≤ 1000, and it slows down after ρ > 1000). When the particle size z > 300 μm (red area), the dispersion degree of the SHAP value increases at the same density, indicating that targeted training data needs to be increased under the condition of large particle size to improve the robustness of the model. This figure verifies the effectiveness of density as a core feature and provides a quantitative basis for the selection of the best model among multiple models.

[0112] Horizontal axis variable ρ:

[0113] Represents the density of particles or fluids (unit: kg / m 3 ³), with a range of 600 - 1200, which is in line with the typical density range for the cyclone to handle liquid-solid / gas-liquid two-phase flows (such as the liquid-solid two-phase flow scenario in claim 2).

[0114] Vertical axis SHAP value:

[0115] Quantifies the contribution intensity of ρ to the prediction result of particle rotational speed. The higher the SHAP value, the greater the impact of density change on the prediction result. The SHAP value in the figure reaches 35000, indicating that ρ is the core sensitive parameter of the model.

[0116] Color bar variable z:

[0117] Presumably the particle size (unit: μm) or turbulent kinetic energy (unit: m 2 ² / s 2 ³), which is consistent with the particle size range (1 - 500 μm) in claim 2 and the data acquisition requirements of turbulent kinetic energy in claim 1.

[0118] Color mapping logic:

[0119] Blue (low z value): Small particle size or low turbulent kinetic energy region, and the influence of ρ on the prediction is relatively stable

[0120] Red (high z value): Large particle size or high turbulent kinetic energy region, and there is a non-linear relationship between ρ and the SHAP value (possibly corresponding to the data compensation mechanism for extreme working conditions in claim 7).

[0121] Feature screening verification:

[0122] Proves that ρ (density) is the key physical quantity for predicting particle rotational speed, forming a closed loop with "weighted fusion of physical quantities such as Coriolis force and vorticity" in claim 2, and guiding feature engineering to preferentially retain variables with high SHAP values.

[0123] Model decision interpretability:

[0124] Please note that the [FIGURE NUMBER] in the translation of ID=4 should be replaced with the actual figure number. Also, the units in the translation are presented in a more standard way for English scientific writing. For example, "kg / m³" for density unit and "m² / s³" for turbulent kinetic energy unit. The specific content may need to be adjusted according to the actual patent context.Meet the requirements of "physical constraint embedding" in the patent (such as using SHAP to explain feature contributions as required in the supplementary description), and reveal that the model conforms to the laws of fluid mechanics (such as an increase in density leading to an increase in centrifugal force, thereby affecting particle rotation).

[0125] Analysis of operating condition adaptability:

[0126] The color mapping reflects the differences in density effects under different particle sizes / turbulent kinetic energy conditions, providing an optimization direction for the "online evolution mechanism" in claim 9 (such as increasing sample training in high z-value regions).

[0127] Reference Figure 4 As shown: Quantify the contribution degrees of the cyclone structure parameters (d = 1.0 / D = 75.0) and operating condition parameters (p = 1200.0 / h = 50.25) to the model prediction result (f(x) = 4310.86) through the SHAP method. The red bars indicate that an increase in the parameter leads to an increase in the predicted value (such as an increase in the inlet diameter d enhancing the particle centrifugal effect), and the blue bars indicate that an increase in the parameter leads to a decrease in the predicted value (such as an increase in the height h extending the flow path and weakening rotation). This figure verifies the multi-parameter coupling influence mechanism and provides an interpretable basis for selecting the best among multiple models.

[0128] Reference Figure 5 As shown: Demonstrate the global influence law of parameters such as z (particle size), ρ (density), d (inlet diameter), etc. on the model output f(x) through the SHAP heatmap. The red area (SHAP>0) indicates that an increase in the parameter leads to an increase in the predicted value of the particle rotation speed (such as a significant positive contribution when ρ = 1200 kg / m 3 ), and the blue area (SHAP<0) indicates that an increase in the parameter leads to a decrease in the predicted value (such as an increase in the height h weakening the centrifugal effect). This figure verifies the complexity of the multi-feature coupling effect, provides a visual basis for selecting the best among multiple models, and guides the setting of the priority of feature weighted fusion.

[0129] Reference Figure 6 As shown: Show the average absolute SHAP values (0 - 1750) of parameters such as z (particle size), p (density), h (height), etc. on the model output through a horizontal bar chart. The SHAP value of z is the largest (1750), indicating that the particle size is the core sensitive parameter for predicting the particle rotation speed; the influence of D (cyclone diameter) is the smallest (close to 0), and feature dimensionality reduction can be combined with claim 3. This figure verifies the rationality of feature weighted fusion and provides a basis for importance ranking for selecting the best among multiple models.

[0130] Reference Figure 7As shown in the figure, the SHAP value distribution of parameters such as z (particle size), ρ (density), and h (height) on the model output is shown through a scatter plot. The horizontal axis represents the SHAP value range from -2500 to 17500. Most of the red points (high feature values) are located in the positive interval (e.g., SHAP > 10000 when ρ = 1200 kg / m 3 ), indicating that high density significantly promotes particle rotation; most of the blue points (low feature values) are located in the negative interval (e.g., SHAP ≈ -1000 when h = 50), indicating that low height inhibits rotation. This figure verifies the rationality of medium feature weighted fusion and provides a directional criterion for multi-model optimization.

[0131] Reference Figure 8 As shown in the figure: The positive and negative contributions of features such as z (particle size = 75 μm), p (density = 1200 kg / m 3 ) to the model prediction value (f(x) = 4455.593) are quantified through a red-blue bar chart. Among them, the positive contribution of z is the largest (+2773.34), indicating that increasing the particle size significantly improves the particle rotation speed; the negative contribution of p is the strongest (-1293.35), indicating that high-density fluid inhibits the rotation effect. This figure verifies the rationality of the feature fusion weight and provides an interpretable criterion for multi-model optimization (preferably select a z / p sensitive model).

[0132] Example: In the production process of a certain chemical enterprise, the hydrocyclone is a key separation device, and its separation efficiency has a crucial impact on product quality and production efficiency. The rotation speed of particles in the hydrocyclone is a key factor affecting the separation effect, but traditional measurement methods are difficult to obtain this parameter in real time and accurately. Therefore, a data-driven rapid prediction method for particle rotation in hydrocyclones is introduced.

[0133] Implementation steps: Data collection

[0134] For data collection, the Fluent-EDEM coupled simulation platform is selected, which adopts the coupling framework of the Euler multiphase flow model and the discrete element method (DEM). In the simulation settings:

[0135] Fluid phase solution:

[0136] Turbulence model: The Reynolds stress model (RSM) is selected to fully solve the transport equations of 7 key terms including the turbulent diffusion term (DT,ij), molecular diffusion term (DL,ij), shear stress generation term (Pij), and pressure strain term (φij);

[0137] Pressure-velocity coupling: The SIMPLE algorithm is adopted

[0138] Inlet condition: Velocity-inlet (corresponding to the inlet gas volume range of 40 - 70 m 3 / h);

[0139] Outlet conditions: both the overflow outlet and the underflow outlet are set to Outflow, and the split ratio adjustment range is 0 - 100%;

[0140] Particle phase solution:

[0141] Material properties: POM plastic particles, density 1200 kg / m 3 , three particle size specifications (1 / 3 / 5 mm);

[0142] Injection method: uniformly generated through the EDEM particle factory at the inlet section;

[0143] Contact model: The Hertz - Mindlin (NoSlip) model is adopted, and the parameter settings are as follows: the collision restitution coefficient between particles is 0.35, the static friction coefficient is 0.55, and the rolling friction coefficient is 0.08; the collision restitution coefficient between particles and the wall is 0.5, the static friction coefficient is 0.3, and the rolling friction coefficient is 0.05;

[0144] Surface treatment: One - sided black spraying is carried out on the white particles to form high - contrast tracer features;

[0145] Mesh generation:

[0146] Unstructured meshes are generated using ICEM software, with a total number of meshes being 86,678 (verified by mesh independence, the maximum deviation of the tangential velocity at the column - cone junction is <5%);

[0147] Local refinement is carried out in the key areas (wall boundary layer, tangential inlet, column - cone junction area);

[0148] Mesh independence verification: The maximum deviation of the tangential velocity for three mesh numbers (86,678 / 173,694 / 285,932) is <5%;

[0149] Coupling settings:

[0150] Time step: Fluent (5×10 -4 s) and EDEM (2×10 -5 s) maintain a 20 - 100 - fold relationship;

[0151] Data exchange: Gas - solid two - way momentum transfer is achieved through the coupling interface, and the particle motion information is fed back to the fluid phase in the form of a momentum source term;

[0152] Total calculation duration: 4 s (reaching the gas - solid two - phase dynamic equilibrium).

[0153] When meshing the cyclone, unstructured grid encryption is emphasized for the wall, inlet, and separation zone. After repeated debugging, it is ensured that the grid independence verification passes, that is, the grid independence verification: the maximum deviation of the tangential velocity for three grid numbers (86,678 / 173,694 / 285,932) is <5%.

[0154] The particle parameters are set according to the actual material properties. The particles are monodisperse POM plastic particles, with three groups of particle sizes (1 / 3 / 5 mm) and a density of 1200 kg / m 3 , and are injected in a uniform distribution at the inlet.

[0155] To obtain high-precision experimental data for calibration, the motion of tracer particles is captured by high-speed photography (frame rate 10 kHz) for verifying the particle rotation trajectory. Using the experimental data as the observed values and the CFD results as the predicted values, a precise and reliable mixed dataset is constructed.

[0156] Feature Extraction and Preprocessing

[0157] Feature engineering processing is performed on the large amount of collected data. First, the PCA dimensionality reduction technique is used to retain the principal components with a cumulative variance contribution rate exceeding 95% for the original complex flow field data, such as velocity vector fields with up to 1000 dimensions, and successfully compress them to 50 dimensions, greatly reducing the complexity of data processing. At the same time, a 3-layer autoencoder (input - 512 - 256 - 50) is designed, using the ReLU activation function, and the decoder is a symmetric structure, with the reconstruction mean square error as the loss function to further explore the deep features of the data.

[0158] The wavelet transform (Daubechies-4 basis function) is used to filter the noise of the data, removing high-frequency noise interference. Then, Min-Max normalization is used to unify the data to the [0,1] interval. For skewed distribution data (such as pressure gradient), the Box-Cox transform is used for optimization processing, and finally a standardized high-quality dataset is constructed.

[0159] Machine Learning Model Construction

[0160] Based on the ensemble learning strategy, a hybrid prediction architecture composed of six types of models, namely backpropagation neural network (BP), CATBoost, random forest (RF), AdaBoost, XGBoost, and support vector machine (SVM), is constructed. The model input is a 50-dimensional spatio-temporal feature sequence, and the output is the scalar value of the particle rotation speed. The specific architecture design is as follows:

[0161] BP network: 3-layer fully connected structure (input layer 50 → hidden layer 64 → output layer 1), ReLU activation, L2 regularization (λ = 0.001);

[0162] CATBoost: Set the tree depth to 8, the number of iterations to 1000, the learning rate to 0.03, and the L2 regularization coefficient to 3;

[0163] Random Forest (RF): The number of decision trees is 200, the maximum depth is 10, and the minimum number of samples in a leaf node is 5;

[0164] AdaBoost: The base learner is a decision tree with a depth of 3, the number of iterations is 200, and the learning rate is 0.1;

[0165] XGBoost: The maximum tree depth is 6, the learning rate is 0.05, the subsampling rate is 0.8, and the regularization parameter α = 1;

[0166] SVM: Radial basis kernel function (RBF), the penalty coefficient C = 10, and the kernel coefficient γ = 0.01;

[0167] Adopt a dynamic weight integration strategy, and allocate model weights based on SHAP value analysis of the validation set: BP(0.15), CATBoost(0.25), RF(0.2), AdaBoost(0.1), XGBoost(0.25), SVM(0.05).

[0168] For model hyperparameter optimization, set the joint search space for multiple models:

[0169] Learning rate: Log range [1e - 4, 1e - 2] (BP / XGBoost / CATBoost);

[0170] Tree depth: 4 - 12 (CATBoost / XGBoost / RF / AdaBoost);

[0171] Regularization coefficient: 0.1 - 10 (SVM / BP / XGBoost);

[0172] Apply the Bayesian optimization algorithm, rely on the Gaussian process surrogate model, iterate 50 times, and select the parameter combination with the largest expected improvement (EI) to ensure that the model has excellent generalization performance.

[0173] Model training and optimization

[0174] Single - model pre - training:

[0175] Tree models (CATBoost / XGBoost / RF / AdaBoost) use the full training set;

[0176] The BP network adopts batch training (batch size 64) and the Adam optimizer (initial learning rate 0.001);

[0177] SVM uses the LIBSVM library for kernel function mapping training;

[0178] Integrated fine-tuning:

[0179] Freeze the parameters of a single model and optimize the integrated weights based on the validation set MAE;

[0180] Use the weighted average method to fuse the outputs of six models, and optimize the weights through grid search;

[0181] To effectively suppress the overfitting phenomenon, introduce a differential regularization strategy:

[0182] BP network: L2 regularization (λ = 0.001) + Dropout(0.2);

[0183] Tree model: Maximum depth constraint + Subsampling rate control;

[0184] SVM: L2 regularization term (C = 10);

[0185] Closely monitor the validation set loss. If the loss of any model does not decrease for 10 consecutive epochs, terminate its training. After multiple rounds of training and optimization, until the model performance meets the convergence standard evaluated by the loss function (mean square error).

[0186] Actual application verification

[0187] During the model deployment stage, realize model lightweighting with the help of an embedded system. Use TensorRT quantization (FP16) for the BP network, and perform pruning optimization on the tree model. The overall model volume is compressed to 30% of the original size, successfully reducing the model volume to 30% of the original size, and it is deployed on the Jetson Xavier NX hardware platform to ensure that the inference latency is controlled within 50ms, meeting the industrial real-time requirements.

[0188] Realize the linkage control with the cyclone control unit. Through the Modbus protocol, receive data from sensors (pressure, flow meter, etc.) every 200ms, and the model outputs the predicted value of the particle rotation speed in real time. Once the predicted rotation speed deviates from the set value by ±15%, immediately trigger the PID controller to adjust the cyclone inlet flow rate, and the adjustment amplitude is accurately calculated according to the formula ΔQ = K_p×e(t) to ensure that the cyclone always maintains an efficient separation state.

[0189] In the error analysis section, set strict indicators: the mean absolute error (MAE) should be less than 5 rad / s, and the coefficient of determination R 2 should be greater than 0.92. If the prediction error exceeds the threshold for 5 consecutive times, the system automatically switches to the CFD online simulation mode, and immediately triggers an alarm to notify the operation and maintenance personnel to check for abnormalities, ensuring the safety and stability of the production process.

[0190] Through the above embodiments, the feasibility, high efficiency and great application value of the data-driven rapid prediction method for the particle self-rotation of cyclones in industrial actual scenarios are fully demonstrated, and it is expected to be popularized in more industries involving the application of cyclones.

[0191] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A data-driven rapid prediction method for cyclone particle rotation, characterized in that: It includes the following steps: S1: Data acquisition: Through computational fluid dynamics (CFD) simulation tools, numerical simulation of the particle motion state in the cyclone is performed to obtain multi-dimensional data including flow field velocity distribution, particle trajectory, pressure gradient and turbulent kinetic energy; S2: Feature extraction and preprocessing: Perform feature engineering processing on the data in step S1, including but not limited to dimension normalization, noise filtering, extraction of key physical quantities, and constructing a standardized data set, where the key physical quantities include but are not limited to centrifugal acceleration and local shear stress; S3: Machine learning model construction: Based on the integration of six models including BP, CATBoost, RF, AdaBoost, XGBoost, and SVM, the prediction model architecture is designed, and the model generalization ability is improved through hyperparameter optimization; S4: Model training and optimization: adopt batch training strategy, combine cross-validation technology to dynamically adjust model parameters, and evaluate model performance through loss function until convergence, where the loss function uses mean square error; S5: Practical application verification: Deploy the trained model to an industrial cyclone device, input sensor or simulation data in real time, output the particle rotation speed prediction results, and perform error comparison analysis with the measured data.

2. The data-driven rapid prediction method for cyclone particle rotation according to claim 1 is characterized by: The CFD simulation tool in step S1 adopts the Euler-Lagrangian coupling model, and the specific implementation includes: The particle phase is modeled using the discrete element method (DEM). The particle size is 1 mm, 3 mm, and 5 mm, and the density is 1200 kg / m 3 ; The Reynolds stress model (RSM) is used to describe the anisotropic turbulence in the fluid phase; The Hertz-Mindlin (NoSlip) contact model is used for gas-solid two-phase coupling; Set the particle-wall collision recovery coefficient to 0.5 and the static friction coefficient to 0.3; The inlet boundary condition is velocity inlet, and the outlet uses free outflow boundary; The time step is set to Fluent 5×10 -4 s and EDEM2×10 -5 s coupling ratio.

3. The data-driven rapid prediction method for cyclone particle rotation according to claim 1 is characterized by: The feature extraction of S2 includes: Dimensionality reduction of high-dimensional data through principal component analysis (PCA) or autoencoders; The operating parameters and structural parameters affecting the particle rotation speed are weightedly fused.

4. The data-driven rapid prediction method for cyclone particle rotation according to claim 1 is characterized by: In the model construction of S3, a hybrid architecture of six types of machine learning models, namely BP neural network, CATBoost, RF, AdaBoost, XGBoost, and SVM, is adopted.

5. The data-driven rapid prediction method for cyclone particle rotation according to claim 1 is characterized by: During the optimization process of S4, a Bayesian optimization algorithm is used to perform a global search for hyperparameters, and an early stopping mechanism is introduced to prevent overfitting.

6. The data-driven rapid prediction method for cyclone particle rotation according to claim 1 is characterized by: In the actual application of the S5, the model is lightweight deployed through the embedded system, and is linked with the cyclone control unit to adjust the separation efficiency parameters in real time.

7. The data-driven rapid prediction method for cyclone particle rotation according to claim 1 is characterized by: In S1, the CFD simulation data and the experimental data are jointly calibrated through a dynamic weight fusion strategy, specifically including: Verify the time-space synchronization of experimental data (PIV measurement, high-speed camera trajectory tracking) and eliminate abnormal sampling points; The Kalman filter algorithm is used to fuse the velocity distribution of the CFD-simulated flow field with the experimental measurement value, and the confidence weight is adjusted dynamically; Construct a hybrid dataset of CFD simulation data and experimental data.

Images