An ejector intelligent prediction optimization method and system

Abstract

This invention discloses an intelligent prediction and optimization method and system for ejectors. The method includes: constructing a multi-dimensional data input system by integrating historical experimental data, numerical simulation results, and real-time operating condition data; extracting key features of multi-physics fields by combining data mining and machine learning techniques; modeling the nonlinear relationship between flow characteristics and design parameters using deep learning algorithms to achieve accurate prediction and dynamic adaptive optimization of ejector performance; and continuously correcting the model accuracy through a closed-loop experimental feedback mechanism. A common technology platform is built based on this method, possessing functions such as multi-dimensional data integration and analysis, intelligent optimization algorithm integration, and experimental feedback iteration, supporting the input and optimization of different operating conditions, fluid media, and design parameters. This invention solves the problems of long cycle time, low accuracy, and poor adaptability to operating conditions in traditional design methods.

Description

Technical Field

[0001] This invention relates to the field of fluid machinery design and intelligent optimization technology, specifically to an intelligent predictive optimization method and system for ejectors. Background Technology

[0002] As a type of fluid machinery without moving parts, ejectors have been widely used in aerospace high-altitude simulation, energy and power fluid transportation, and chemical process intensification due to their advantages such as simple structure, high reliability, and adaptability to extreme working conditions. However, their performance is affected by multiple factors such as nozzle structure, mixing chamber size, and operating parameters. The internal flow field involves complex physical processes such as supersonic flow, shock wave interference, and multiphase mixing.

[0003] Traditional design methods have significant shortcomings: they rely on empirical trial and error or full factorial experiments, requiring extensive numerical simulations and physical experiments, with development cycles typically exceeding 6 months; they often employ single-parameter optimization, failing to adequately consider the coupling effects between parameters, making them prone to getting trapped in local optima, resulting in generally low improvements in ejector efficiency; they lack the ability to adaptively adjust to dynamically changing operating conditions; and design methods for specific ejectors are difficult to transfer to other fluid equipment, resulting in poor cross-domain technology reusability.

[0004] Therefore, there is an urgent need for an ejector design method that integrates multi-source data, intelligent algorithms and closed-loop verification, and to build a common technology platform with cross-domain adaptability to achieve efficient, accurate and adaptive ejector design. Summary of the Invention

[0005] In view of this, the present invention proposes an intelligent prediction and optimization method and system for ejectors. By fusing experimental data and numerical simulation data, multi-physics coupling modeling, intelligent algorithm training and adaptive optimization, global optimization and platform integration, it can achieve accurate prediction of ejector performance and global optimal configuration of design parameters. At the same time, it supports the reuse of technologies from multiple fields, reduces R&D costs and shortens the R&D cycle.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention proposes an intelligent prediction and optimization method for ejectors, comprising:

[0008] Multiple physical tests were conducted on the target ejector under typical operating conditions to obtain real test data on the ejector's performance.

[0009] A multiphysics coupled numerical model was constructed, and high-precision numerical simulations were performed under the same typical working conditions to obtain simulation data containing phase transition behavior and multi-medium mixing characteristics.

[0010] Experimental data and simulation data are fused to form a joint training dataset. A hybrid modeling strategy of physics-driven and data-driven approaches is adopted to train and establish an intelligent prediction model between ejector structural parameters, operating parameters and performance indicators.

[0011] Based on the prediction error or uncertainty of the intelligent prediction model, an adaptive optimization mechanism is triggered to incrementally update or retrain the model, continuously improving the prediction accuracy.

[0012] Using the updated intelligent prediction model as a proxy model, a global optimization algorithm is used to search for the optimal solution of the design parameters, and the solution is verified by high-fidelity numerical simulation or physical experiment, forming a closed-loop iteration of model optimization and parameter optimization.

[0013] A common technology platform is built to provide standardized data interfaces to support users in uploading experimental data, automatically triggering model iteration and updates, and enabling the reuse of technologies for the design optimization of ejectors and similar fluid equipment.

[0014] Preferably, physical tests are conducted on existing ejectors under multiple typical operating conditions, including: designing an operating condition matrix covering different initial pressure and temperature combinations of primary and secondary flows; using a closed-loop gas supply system and high-precision measuring equipment to synchronously collect pressure, temperature, and flow parameters; and ensuring the reliability and consistency of test data through repeated testing and outlier removal.

[0015] Preferably, a multi-physics coupled numerical model is constructed, including: using a component transport model to simulate the multi-medium mixing process; using a real gas model to describe the changes in the working fluid's physical properties with temperature and pressure; using a phase change model to capture local condensation behavior; using a turbulence model to characterize the flow pulse characteristics; and achieving high-precision numerical reconstruction of complex physical fields through model parameter adaptation and convergence criterion settings.

[0016] Preferably, a hybrid modeling strategy combining physical and data-driven approaches is used to train the intelligent prediction model, including: constructing a rapid screening model and performing preliminary performance evaluation on a large parameter space based on a gradient boosting ensemble mechanism; constructing a refined prediction model and using a multi-layer fully connected neural network to perform high-precision regression prediction on key parameter regions; and using a Bayesian surrogate model to output prediction uncertainty and quantify the model confidence level.

[0017] Preferably, the training and establishment of an intelligent prediction model includes: cleaning, consistency checking, outlier handling and missing value completion of experimental and simulated data; dividing the data into training, validation and test sets; constructing original features and physical derived features as model inputs; using weighted loss functions and regularization methods to suppress overfitting; and evaluating the prediction accuracy of the model under extreme low and high temperature conditions through independent validation.

[0018] Preferably, the adaptive optimization mechanism triggered by prediction error or uncertainty includes: comparing the model output with newly added experimental data or high-fidelity simulation results to calculate the prediction deviation and uncertainty level; determining that the model's generalization ability has decreased when the error exceeds a preset threshold or the uncertainty is higher than a set upper limit; automatically merging the corresponding working condition data into the training set for incremental updates or retraining; evaluating the performance of the updated model through a validation dataset, and replacing the original model if the requirements are met.

[0019] Preferably, a global optimization algorithm is used to search for the optimal solution of the design parameters, including: using an intelligent prediction model as a surrogate model to replace high-fidelity calculation and establishing a fast mapping between design parameters and performance indicators; using a Bayesian optimization algorithm to perform a global search under constraints, and combining the performance prediction values ​​and uncertainty information output by the model to screen candidate solutions; performing high-fidelity numerical simulations or physical experiments to verify the candidate parameters, and calculating the relative error between the predicted values ​​and the true values; when the error exceeds a threshold, feeding the verification data back to the model for correction, restarting the parameter search, and forming a closed-loop iteration until a stable optimal solution is obtained.

[0020] Preferably, a common technology platform is constructed, including: providing a standardized data interface to support users in uploading experimental test data and numerical simulation results of ejectors and similar fluid equipment; automatically parsing, formatting, normalizing, and evaluating the quality of uploaded data; uniformly mapping data into a standardized form to support one-click model training; integrating a model management module to store, identify, and call different versions of fast prediction models, fine prediction models, and uncertainty models; and deploying a model application service module to quickly output performance prediction results and their uncertainty information based on user-input structural design parameters and operating condition parameters, and calling the trained model.

[0021] To achieve the above objectives, the present invention also provides an intelligent predictive optimization system for ejectors, comprising:

[0022] The data acquisition module is used to perform physical experiments and obtain real experimental data;

[0023] The multiphysics simulation module is used to build coupled numerical models to obtain simulation data;

[0024] The intelligent prediction module is used to integrate experimental and simulation data and employs a hybrid modeling strategy to train a rapid screening model, a refined prediction model, and a Bayesian surrogate model.

[0025] The adaptive optimization module is used to monitor prediction errors and uncertainties and trigger an adaptive update mechanism.

[0026] The global optimization module is used to perform global optimization by using an intelligent prediction model as an agent, and forms a closed-loop iteration through high-fidelity verification.

[0027] The common platform module provides standardized data interfaces, model training management, and service deployment, supporting cross-equipment technology reuse.

[0028] Preferably, the data acquisition module includes a closed-loop gas supply system, a pressure sensor, a temperature sensor, a flow meter, and a data acquisition and analysis unit; the multiphysics simulation module includes a component transport solver, a real gas property library, a phase change model library, and a turbulence model library; the common platform module includes a standardized data upload interface, a data parsing and verification unit, a normalization processing unit, a model training trigger unit, a model version management unit, and a prediction service deployment unit.

[0029] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art.

[0030] 1. By combining multiple models and adapting parameters, the phase transition law and multi-medium mixing characteristics inside the ejector are accurately captured, solving the problem of large deviation between traditional simulation and reality;

[0031] 2. By integrating experimental and simulated data and using deep learning algorithms to construct a mapping model, performance can be predicted quickly, balancing accuracy and efficiency;

[0032] 3. By combining Bayesian optimization and multiphysics analysis, model parameters are dynamically adjusted to continuously improve prediction accuracy;

[0033] 4. Employ a global optimization algorithm to overcome local optima and achieve optimal configuration of design parameters;

[0034] 5. Build a common technology platform to support user data uploading and model iteration, realize the reuse of technologies in multiple fields, and reduce R&D costs.

[0035] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0036] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:

[0037] Figure 1 This is a flowchart illustrating the steps of an intelligent prediction and optimization method for ejectors in an embodiment of the present invention.

[0038] Figure 2 This describes the relationship and detailed process between the steps in the embodiments of the present invention. Detailed Implementation

[0039] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0040] To achieve the above objectives, the present invention provides the following technical solution:

[0041] This invention provides an intelligent prediction and optimization method and system for ejectors, offering the following technical solution:

[0042] Typical operating condition test data acquisition: Select the existing ejector as the test object, design multiple sets of typical operating conditions to carry out physical tests, and the operating condition variables include key parameters such as the initial temperature and pressure of the primary flow and secondary flow. The real performance data of the ejector is obtained through experimental measurement, which provides a reliable experimental basis for subsequent model training.

[0043] High-precision multiphysics numerical simulation: Addressing phase change phenomena (such as gas-liquid conversion between media with different boiling points) and multi-media mixing behavior within the ejector, this approach abandons traditional numerical simulation methods that neglect phase change effects. Instead, it employs a combination of component transport models, real gas models, phase change models, and turbulence models. By adjusting the model parameters for adaptability, a high-precision numerical simulation model is established. Simulation calculations are performed under multiple typical operating conditions identical to those in physical experiments to obtain high-precision simulation data encompassing phase change laws, media mixing characteristics, and turbulence effects.

[0044] Intelligent prediction model training: Integrate experimental and numerical simulation data to construct a comprehensive dataset; use deep learning algorithms to extract features from the dataset, explore the intrinsic relationship between ejector design parameters, operating parameters and performance indicators, and establish an accurate mapping model between the three through model training and iterative optimization, so as to achieve rapid prediction of ejector performance under different parameter combinations and solve the problem of long time consumption in traditional numerical simulation.

[0045] Adaptive optimization mechanism establishment: Based on the output results of the constructed intelligent prediction model, a dynamic adaptive optimization mechanism is established; Bayesian optimization and other algorithms are used to iteratively correct the model parameters, and the prediction accuracy of the model for complex flow phenomena is continuously calibrated by combining the results of multiphysics coupling analysis, so as to obtain a high-precision intelligent prediction model and ensure the reliability of the prediction results.

[0046] Global optimal search for design parameters: Based on the obtained high-precision intelligent prediction model, a global optimization algorithm is used to perform a global search for the key design parameters of the ejector (including but not limited to the nozzle throat diameter, the length-to-diameter ratio of the mixing chamber, and the diffuser angle). By breaking through the limitations of local optimal solutions, the global optimal configuration of the ejector performance is achieved, thereby improving the working efficiency and applicability of the ejector.

[0047] Common Technology Platform Construction: An experimental data upload interface has been developed, allowing users to upload their own ejector physics experiment results to the platform and compare them with the platform's prediction results. When the comparison deviation exceeds a preset threshold, the platform automatically triggers a model iteration update mechanism to continuously optimize model accuracy. Simultaneously, the platform integrates multi-condition data management and multi-domain algorithm adaptation functions, constructing a common technology platform applicable to multiple industrial fields such as aerospace and energy, enabling the reuse and cross-domain promotion of ejector and similar fluid equipment design optimization technologies.

[0048] like Figure 2 As shown in the figure, this invention provides a schematic diagram of the intelligent prediction and optimization method for ejectors and the construction process of a common technology platform.

[0049] The following is in conjunction with the appendix Figure 1 as well as Figure 2 This paper details the implementation process, parameter settings, and verification procedures of the intelligent prediction and optimization method and common technology platform for ejectors of the present invention. This embodiment takes a supersonic ejector in the aerospace field as the research object, aiming to optimize its ejection ratio and pressure recovery coefficient to achieve efficient fluid transport under extreme conditions. Specifically, the intelligent prediction and optimization method and common technology platform for ejectors includes the following steps:

[0050] Step 1: Select an existing supersonic ejector as the test object. This ejector is mainly used in the fluid transport system of aerospace engine high-altitude simulation test bench. The core operating conditions cover two scenarios: ground test (normal pressure, normal temperature) and high-altitude simulation (low pressure, low temperature). Based on actual application requirements, multiple typical operating conditions are designed. The operating condition variables focus on the two core parameters of initial pressure and temperature of the primary and secondary flows. The primary flow nitrogen pressure is 0.8-2.5 MPa and the temperature is 300-500 K; the secondary flow hydrogen pressure is 0.1-0.3 MPa and the temperature is 50-100 K. The outlet is room temperature air at atmospheric pressure. The specific value ranges are determined by surveying industry application standards and actual test data.

[0051] Step 1.1: The experiment was conducted on a national-level high-precision fluid machinery testing rig (CNAS certified). The testing rig is equipped with a closed-loop air supply system, a temperature control system, and a data acquisition and analysis system. The selection and calibration of the core measuring equipment strictly followed the preset standards.

[0052] Step 1.2: Calibrate all measuring equipment before the test: calibrate the pressure transmitter with a piston pressure gauge, the temperature sensor with a constant temperature bath, and the flow meter with a standard flow calibration device. The calibration certificate must be valid, and a zero-point calibration must be performed after every two sets of working conditions during the test to eliminate equipment drift error.

[0053] Step 1.3: Pressure Measurement: A Rosemount 3051CG intelligent pressure transmitter is used, with a measurement range of 0-3.0 MPa and an accuracy of ±0.1%FS. It is installed at five key sections: primary flow inlet, secondary flow inlet, vacuum chamber inlet, and outlet, with a sampling frequency of 10 Hz. Temperature Measurement: A rhodium-iron resistance thermometer (measurement range 0.5~273K) and an armored platinum resistance temperature sensor (measurement range 200~600K) are used, with their installation positions corresponding to the pressure sensors, and the fluid temperature at each section is collected synchronously.

[0054] Step 1.4: Flow measurement: The primary flow uses a Coriolis mass flow meter (accuracy ±0.15%), with a measurement range of 0-50 kg / h; the secondary flow uses a vortex flow meter (accuracy ±1%), with a measurement range of 0-20 kg / h. Both are installed in the straight section before the fluid inlet (straight pipe length ≥ 10 times the pipe diameter) to avoid flow field disturbances affecting measurement accuracy.

[0055] Step 1.5: Repeat the test 3 times for each operating condition, with a test interval of ≥10 minutes, to ensure that the ejector flow field completely recovers to its initial state and avoids the influence of residual heat and pressure from the previous test. After each test, first remove outlier data points: use the 3σ criterion to screen, marking data that deviate from the mean by more than 3 times the standard deviation as outliers, and analyze the causes of outliers (such as instantaneous sensor interference, gas source pressure fluctuations). If it is due to equipment failure, retest; if it is due to random interference, remove outliers and calculate the average of the remaining data. The results of the three repeated tests must meet the consistency requirements: under the same operating condition, the coefficient of variation of the ejector ratio (secondary flow rate / primary flow rate) for the three test results should be ≤2%, and the coefficient of variation of the pressure recovery coefficient (diffuser outlet pressure / mixing chamber inlet pressure) for the three test results should be ≤3%. Otherwise, increase the number of tests to 5, and calculate the average of the 4 sets of data with the coefficient of variation meeting the requirements. Finally, a set of measured datasets for multiple operating conditions is formed, including key indicators such as primary flow rate, secondary flow rate, pressure / temperature at each section, and ejector ratio, providing a reliable measured basis for subsequent model training.

[0056] Step 2: In view of the characteristics of multi-medium mixing and local condensation phase change of nitrogen and cryogenic hydrogen in the ejector of this embodiment, the traditional single turbulence model simulation method that ignores the phase change effect is abandoned, and the complex physical field is accurately captured by multi-model coupling.

[0057] Step 2.1: Component Transport Model: Set the physical properties of the two gases. Density varies according to the ideal gas law, specific heat is determined according to the mixing law, thermal conductivity and viscosity are determined according to the mass-weighted mixing law, and mass diffusivity is given as a constant of 2.88e⁻⁵ m² / s. Accurately simulate the mixing process and concentration distribution of the two media in the mixing chamber;

[0058] In detail, the density of the mixture is given by the following formula:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] in, These are absolute, operational, and relative pressures, respectively. These are the gas mixture constant, the universal gas constant, and the molar mass of the mixture, respectively. These are the mass fraction and molar mass of component i, respectively.

[0065] The specific heat of the mixture is given by the following formula:

[0066]

[0067] These are the specific heat of the mixture, the specific heat of component i in the mixture, and the mass fraction of component i, respectively.

[0068] The thermal conductivity and viscosity of the mixture are given by the following formula:

[0069]

[0070]

[0071] , These are the thermal conductivity and viscosity of component i, respectively.

[0072] The specific heat and thermal conductivity of each component were obtained using discrete polynomials, and the viscosity was obtained using the Sutherland formula.

[0073] In detail, the viscosity of each component is given by the following formula:

[0074]

[0075] in, These are the reference temperatures. The dynamic viscosity, temperature, reference temperature, and Sutherland constant (97 K for hydrogen and 107 K for nitrogen) are given.

[0076] In detail, taking nitrogen as an example, its specific heat and thermal conductivity are obtained through the following formulas:

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] Step 2.2: Condensation phase change model: Using the saturation temperature of nitrogen in the mixed fluid as the core criterion, combined with local pressure and component concentration corrections, the spatial range of phase change occurrence is determined.

[0084] Saturation temperature calculation: The relationship between nitrogen saturation temperature and pressure is described based on the Antoine equation, as shown in the following formula:

[0085]

[0086] in, These represent the saturated vapor pressure (Pa) and local fluid temperature (K) of hydrogen at the current pressure, respectively. For Antoine's constant, the values ​​are 15.702, 1064.8, and -43.4.

[0087] Phase change triggering condition: local fluid temperature ( When the hydrogen mass fraction is greater than or equal to the hydrogen mass fraction in the gas phase under saturated conditions (which is derived from the ideal gas law combined with the saturated vapor pressure), nitrogen condensation phase transition is triggered; otherwise, no phase transition occurs.

[0088] Interphase mass transfer model: The source term method is used to introduce phase change mass exchange into the governing equations, and the liquid phase nitrogen generation rate (condensation rate) is defined:

[0089]

[0090] in, The mass generation rate of liquid nitrogen (kg / (m³·s)) is represented by a positive value indicating condensation. The density of the mixed fluid (kg / m³); The density of the mixed fluid under saturated conditions (kg / m³); The phase transition relaxation time (s) reflects the phase transition rate and is taken as 5e-4.

[0091] Perform source term allocation: gaseous nitrogen mass source term (gaseous mass consumption).

[0092]

[0093] The liquid phase nitrogen mass source term is (increase in liquid phase mass):

[0094]

[0095] Further phase change thermodynamic coupling correction is performed: the latent heat released during the condensation process is introduced through the source term of the energy equation, affecting the temperature distribution of the flow field.

[0096]

[0097] In the formula These represent the source terms of the energy equation and the latent heat of nitrogen condensation (which can be corrected with polynomial fitting based on temperature). ).

[0098] Step 2.3: The multi-model coupling process described above is implemented using OpenFOAM. Based on multiple typical operating conditions from physical experiments, the simulation covers key parameter ranges for mainstream / ejector fluid pressure, temperature, and mixing chamber back pressure. The focus is on investigating the phase change characteristics under low / normal / high loads, the influence of temperature and pressure parameters on phase change, and the flow coupling effect under high back pressure. Strict convergence criteria are set for the simulation calculations, and the outlet mass flow rate and residual curves are monitored. When the operating conditions change slightly, the previous convergence result is used as the initial condition, which can significantly improve computational efficiency and stability.

[0099] Step 3: The intelligent prediction model adopts a hybrid modeling strategy of physics-driven and data-driven approaches. Based on high-precision multiphysics numerical simulations and high-quality experimental data, it extracts key variables from the multiphysics field and constructs a machine learning deep learning surrogate model to quickly predict ejector performance indicators (such as ejection ratio, vacuum chamber pressure, etc.). Uncertainty estimation and adaptive update mechanisms are embedded in the prediction model, and Bayesian optimization / global optimizer is used for parameter optimization. This module, as the core algorithm unit of the common platform, is tightly coupled with the data upload interface, model iteration triggering mechanism, and experimental-simulation comparison closed loop.

[0100] Step 3.1: The model input consists of the following aspects:

[0101] Design parameters: Nozzle throat diameter , Mixed chamber length-to-diameter ratio Diffusion section angle , nozzle outlet specific area, etc.;

[0102] Operating parameters: Primary flow pressure Primary flow temperature Secondary flow pressure Secondary flow temperature Export back pressure Fluid component mass fraction, etc.;

[0103] The model output includes the following aspects:

[0104] Performance indicators: ejection ratio, vacuum chamber pressure, outlet pressure;

[0105] Error / uncertainty measures: prediction variance, uncertainty confidence interval, etc. (used for adaptive triggering).

[0106] Step 3.2: Data cleaning: Remove sensor fault data, identify and process outliers using the 3σ principle; perform consistency checks on simulation data and experimental data (time, coordinate system, units); and complete missing values ​​using physical constraint interpolation or interpolation based on similar working conditions.

[0107] After uniformly collecting ejector test data, multiphysics numerical simulation data, and historical engineering data, the source, acquisition time, acquisition method, data version, and corresponding operating conditions of each data sample are first verified. If there are data samples with missing fields, incomplete data records, or unclear acquisition information, the samples are marked, and a decision is made on whether to remove them based on their importance.

[0108] Based on the physical boundary conditions of ejector operation and engineering experience, outliers in the data are identified. Specifically, this includes: data that clearly violates physical laws, such as negative pressure, negative temperature, sudden flow changes, or data exceeding material and structural limits; data that deviates significantly from data of the same operating conditions; time series data exhibiting non-physical oscillations or sudden changes under stable operating conditions; and the 3σ principle: using the mean and standard deviation of the data to define a reasonable data range, with values ​​exceeding this range being judged as outliers.

[0109] For the detected abnormal data, different processing methods are adopted according to their quantity and degree of deviation: for a small number of isolated outliers, interpolation or smoothing correction can be performed using nearby stable data; for data samples with systematic anomalies or large-scale deviations, they are directly removed from the model training set; for data with some uncertainty but still of reference value, their weight is reduced in the model training to reduce their impact on the prediction results.

[0110] In actual experiments and simulations, some sensor signals or calculation results may be missing. The following principles are adopted for handling missing data: For missing data of non-critical variables, estimation or interpolation is used to complete the data using other relevant physical quantities; for missing data corresponding to key performance indicators (such as ejection ratio, pressure recovery coefficient, etc.), it is generally not used as supervised samples in model training; the completed data is labeled so that the model can identify differences in confidence levels during training, thereby improving model robustness.

[0111] To avoid unit discrepancies between different data sources, all data must be uniformly converted to a pre-defined standard unit system, such as the International System of Units (SI). This includes, but is not limited to, physical quantities such as pressure, temperature, length, mass flow rate, and velocity.

[0112] Step 3.3: Divide the dataset into training set: validation set: test set = 70%: 15%: 15% (ensure that all operating conditions are distributed proportionally).

[0113] Step 3.4: Construct the original features and physically derived features. The original features are directly derived from the design and operating parameters (see Section 2). Derived features include local Mach numbers (used to determine whether it is supersonic), etc.

[0114] Step 3.5: Adopt a serial hybrid agent architecture of "fast screening-refined prediction" to achieve a balance between efficiency and accuracy through the collaboration of two models.

[0115] Rapid model selection: Gradient Boosting Machines (GBM) is selected. Relying on its gradient boosting ensemble mechanism, it has the advantages of high training efficiency, strong model interpretability and small sample data adaptation. It can quickly evaluate the ejector performance in a wide range of parameter spaces to achieve efficient preliminary selection.

[0116] A refined prediction model is constructed using a multi-layer fully connected neural network. This network receives normalized feature vector inputs and has 3 to 6 hidden layers, each containing 64 to 512 neurons. The ReLU activation function is used to achieve non-linear mapping, and the output layer uses linear activation to complete the regression task. During training, a weighted mean squared error loss function (WeightedMSE) is employed. The weights are positively correlated with sample confidence and measurement uncertainty. L2 regularization is used to suppress overfitting, enabling high-precision prediction of ejector performance within key parameter regions, thereby improving the reliability of the final design parameters.

[0117] Step 3.6: Uncertainty modeling adopts the Bayesian surrogate Gaussian Process (GP) algorithm. By constructing a surrogate model with probability output capability, the uncertainty characterization of the ejector performance prediction results is given, which is used to reflect the prediction confidence level of the fine prediction model in different parameter regions.

[0118] Step 3.7: Validation Strategy. The reserved test set is only used for the final evaluation to statistically analyze the prediction accuracy under low-temperature and high-temperature conditions.

[0119] Step 4: Based on the output of the constructed intelligent prediction model, establish a dynamic adaptive optimization mechanism.

[0120] Step 4.1: After the intelligent prediction model completes the ejector performance prediction, the model output results are compared and analyzed with newly acquired experimental data or high-fidelity simulation results to calculate the prediction deviation and uncertainty level of the ejection ratio and vacuum chamber pressure. When the prediction error exceeds the preset threshold, or the prediction uncertainty given by the model for a specific working condition or parameter combination is higher than the set upper limit, it is determined that the generalization ability of the current model in the corresponding parameter region has decreased, triggering adaptive optimization conditions to start the model update or supplementary training process.

[0121] Step 4.2: After triggering the adaptive optimization condition, the new experimental data or simulation results under the corresponding working condition are automatically incorporated into the training dataset to incrementally update or retrain the intelligent prediction model. During the update process, the existing model structure and physical constraints are preserved, and only the model parameters are adaptively corrected to improve the prediction accuracy within the parameter space dominated by the newly added data. After the update is complete, the model performance is evaluated using an independent validation dataset. If the error requirements are met, the original model is replaced and used in subsequent parameter prediction and optimization processes.

[0122] Step 5: Using the obtained high-precision intelligent prediction model as the core, a global optimization algorithm is used to perform a global search for the key design parameters of the ejector.

[0123] Step 5.1: After completing the training of the intelligent prediction model and passing the accuracy verification, the intelligent prediction model is used as a proxy model for ejector performance evaluation, replacing high-fidelity numerical simulation or physical experiments to achieve rapid performance prediction of multi-dimensional design parameters. The design parameter vector is...

[0124]

[0125] in Where L is the nozzle throat diameter, D is the mixing chamber length, and D is the mixing chamber diameter. The diffuser angle; the ejector performance index is defined as follows:

[0126]

[0127] This function The results are provided by an intelligent prediction model. Based on this, a global optimization strategy is introduced to quickly search the design space while satisfying geometric constraints, operating condition constraints, and physical constraints. The constraints can be expressed as follows:

[0128] x min <x<x max ,g j (x)<0

[0129] By using the Bayesian optimization algorithm, combined with the performance predictions and uncertainty information output by the surrogate model, the candidate parameters are iteratively updated, and several candidate design schemes with significant performance advantages in the full parameter space are selected. This effectively avoids getting trapped in local optima and significantly reduces the number of high-fidelity computation calls.

[0130] Step 5.2: For the candidate parameter combinations obtained from the global parameter search, further use a high-fidelity numerical simulation model or physical experimental platform to verify their actual performance and obtain the actual ejection performance indicators under the corresponding parameter combinations. The verification results are compared with the predicted values ​​given by the intelligent prediction model. Comparative analysis was performed, and the relative error was calculated:

[0131]

[0132] When the error is less than a preset threshold (e.g., 5%–10%) and all engineering constraints are met, the corresponding parameter combination can be determined as the globally optimal design parameters. If the error exceeds the threshold, the validation data is fed back to the intelligent prediction model as a new sample to correct the model parameters or perform local retraining. The parameter search process is then restarted on the basis of the updated model until a stable and repeatable optimal design solution is obtained, thus forming a closed-loop iterative mechanism for parameter optimization and model updating.

[0133] Step 6: Build a common technology platform based on the above model.

[0134] Step 6.1: Based on the aforementioned intelligent prediction model and global optimization mechanism, a common technology platform for ejector and similar fluid equipment design is constructed to achieve multi-source data access, unified modeling support, and intelligent prediction and optimization services across equipment types. This common technology platform provides standardized data interfaces, supporting users to upload experimental test data and numerical simulation results for ejectors and similar fluid equipment, and automatically parses and verifies the quality of uploaded data. The platform achieves data consistency across experimental conditions and simulation tools by uniformly processing data from different sources and in different formats. The platform uniformly maps data into the following standardized form:

[0135]

[0136] in Indicates structural design parameters, Indicates the corresponding operating condition parameters. This indicates the performance metrics obtained from experiments or simulations.

[0137] Step 6.2: By normalizing the data, identifying outliers, and verifying label integrity, users can start model training with one click. The platform's model integration and management module is used to store, identify, and call different versions of intelligent prediction models, including fast prediction models, fine prediction models, and prediction models with uncertainty output capabilities.

[0138] Step 6.3: The platform's model deployment and application service module supports calling the trained intelligent prediction model based on user-input structural design parameters and operating condition parameters, and quickly outputting ejector performance prediction results and their uncertainty information. In a preferred embodiment, the intelligent prediction model can be trained based on a neural network structure; however, the present invention is not limited to the above implementation method, and the model can also be implemented using other prediction models with nonlinear mapping capabilities. Through this module, the platform can provide a unified data support and computing service foundation for intelligent modeling and optimization of different types of ejectors and related fluid equipment.

[0139] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A method for intelligent prediction and optimization of ejectors, characterized in that, include: Multiple physical tests were conducted on the target ejector under typical operating conditions to obtain real test data on the ejector's performance. A multiphysics coupled numerical model was constructed, and high-precision numerical simulations were performed under the same typical working conditions to obtain simulation data containing phase transition behavior and multi-medium mixing characteristics. Experimental data and simulation data are fused to form a joint training dataset. A hybrid modeling strategy of physics-driven and data-driven approaches is adopted to train and establish an intelligent prediction model between ejector structural parameters, operating parameters and performance indicators. Based on the prediction error or uncertainty of the intelligent prediction model, an adaptive optimization mechanism is triggered to incrementally update or retrain the model, continuously improving the prediction accuracy. Using the updated intelligent prediction model as a proxy model, a global optimization algorithm is used to search for the optimal solution of the design parameters, and the solution is verified by high-fidelity numerical simulation or physical experiment, forming a closed-loop iteration of model optimization and parameter optimization. A common technology platform is built to provide standardized data interfaces to support users in uploading experimental data, automatically triggering model iteration and updates, and enabling the reuse of technologies for the design optimization of ejectors and similar fluid equipment.

2. The intelligent prediction and optimization method for ejectors according to claim 1, characterized in that, The physical tests conducted on the target ejector under multiple typical operating conditions include: designing an operating condition matrix covering different combinations of initial pressure and temperature for primary and secondary flows; using a closed-loop gas supply system and high-precision measuring equipment to synchronously collect pressure, temperature, and flow parameters; and ensuring the reliability and consistency of the test data through repeated testing and outlier removal.

3. The intelligent prediction and optimization method for an ejector according to claim 1, characterized in that, The construction of the multiphysics coupled numerical model includes: using a component transport model to simulate the multi-medium mixing process; using a real gas model to describe the changes in the working fluid's physical properties with temperature and pressure; using a phase change model to capture local condensation behavior; using a turbulence model to characterize the flow pulse characteristics; and achieving high-precision numerical reconstruction of complex physical fields through model parameter adaptation and convergence criterion settings.

4. The intelligent prediction and optimization method for an ejector according to claim 1, characterized in that, The method of training the intelligent prediction model using a hybrid modeling strategy of physical-driven and data-driven approaches includes: constructing a rapid screening model and performing preliminary performance evaluation on a large parameter space based on a gradient boosting ensemble mechanism; constructing a refined prediction model and using a multi-layer fully connected neural network to perform high-precision regression prediction on key parameter regions; and using a Bayesian surrogate model to output prediction uncertainty and quantify the model confidence level.

5. The intelligent prediction and optimization method for an ejector according to claim 4, characterized in that, The training process for establishing an intelligent prediction model includes: cleaning, consistency checking, outlier handling, and missing value completion of experimental and simulated data; dividing the data into training, validation, and test sets; constructing original features and physically derived features as model inputs; using a weighted loss function and regularization method to suppress overfitting; and evaluating the model's prediction accuracy under extreme low and high temperature conditions through independent validation.

6. The intelligent prediction and optimization method for an ejector according to claim 1, characterized in that, The adaptive optimization mechanism triggered by prediction error or uncertainty includes: comparing the model output with newly added experimental data or high-fidelity simulation results to calculate the prediction deviation and uncertainty level; determining that the model's generalization ability has decreased when the error exceeds a preset threshold or the uncertainty is higher than a set upper limit; automatically merging the corresponding working condition data into the training set for incremental updates or retraining; evaluating the performance of the updated model through a validation dataset, and replacing the original model if the requirements are met.

7. The intelligent prediction and optimization method for an ejector according to claim 1, characterized in that, The process of using a global optimization algorithm to search for the optimal solution for design parameters includes: using an intelligent prediction model as a surrogate model to replace high-fidelity calculations and establishing a fast mapping between design parameters and performance indicators; using a Bayesian optimization algorithm to perform a global search under constraints, and combining the performance prediction values ​​and uncertainty information output by the model to screen candidate solutions; performing high-fidelity numerical simulations or physical experiments to verify the candidate parameters, and calculating the relative error between the predicted values ​​and the actual values; when the error exceeds a threshold, feeding the verification data back to the model for correction, restarting the parameter search, and forming a closed-loop iteration until a stable optimal solution is obtained.

8. The intelligent prediction and optimization method for an ejector according to claim 1, characterized in that, The aforementioned common technology platform includes: providing a standardized data interface to support users in uploading experimental test data and numerical simulation results of ejectors and similar fluid equipment; automatically parsing, formatting, normalizing, and evaluating the quality of uploaded data; uniformly mapping data into a standardized form to support one-click model training; integrating a model management module to store, identify, and call different versions of fast prediction models, fine prediction models, and uncertainty models; and deploying a model application service module to quickly output performance prediction results and their uncertainty information based on user-input structural design parameters and operating condition parameters, using the trained model.

9. An intelligent predictive optimization system for ejectors, characterized in that, include: The data acquisition module is used to perform physical experiments and obtain real experimental data; The multiphysics simulation module is used to build coupled numerical models to obtain simulation data; The intelligent prediction module is used to integrate experimental and simulation data and employs a hybrid modeling strategy to train a rapid screening model, a refined prediction model, and a Bayesian surrogate model. The adaptive optimization module is used to monitor prediction errors and uncertainties and trigger an adaptive update mechanism. The global optimization module is used to perform global optimization by using an intelligent prediction model as an agent, and forms a closed-loop iteration through high-fidelity verification. The common platform module provides standardized data interfaces, model training management, and service deployment, supporting cross-equipment technology reuse.

10. The intelligent predictive optimization system for ejectors according to claim 9, characterized in that, The data acquisition module includes a closed-loop gas supply system, a pressure sensor, a temperature sensor, a flow meter, and a data acquisition and analysis unit; the multiphysics simulation module includes a component transport solver, a real gas property library, a phase change model library, and a turbulence model library; the common platform module includes a standardized data upload interface, a data parsing and verification unit, a normalization processing unit, a model training triggering unit, a model version management unit, and a prediction service deployment unit.

Images