A method for predicting flow characteristics in a pipe network of an environmental simulation system

Abstract

The application provides a kind of environmental simulation system pipe network flow characteristic prediction method, belongs to the technical field of aero-engine test bed environmental simulation and fluid flow characteristic prediction.This method explores the evolution mechanism of environmental simulation system;Through the time series data of outlet flow velocity and velocity gradient of pipe network collected by transition state test data acquisition, and deduce the calculation formula and steps of outlet pressure and temperature of different pipe network structure;Feature selection is performed on the preprocessed time series data, the correlation between all selected features and outlet temperature correction coefficient and outlet pressure correction coefficient is calculated to obtain key features;Combined with the transition state test data for correction, and through machine learning algorithm to establish the correction coefficient prediction model of outlet pressure and outlet temperature;The pipe network flow characteristics are predicted by using the correction coefficient prediction model.The application proposes a mixed pipe network flow characteristic prediction method, which ensures basic adaptability through mechanism modeling and improves accuracy through data-driven correction.

Description

Technical Field

[0001] This invention belongs to the field of environmental simulation and fluid flow characteristic prediction technology of aero-engine test bench, and particularly relates to a method for predicting the flow characteristics of pipeline network in an environmental simulation system. Background Technology

[0002] Pipeline flow characteristic prediction is a core technology in environmental simulation systems for aero-engine test benches. Its prediction accuracy, adaptability, and real-time efficiency directly determine the reliability and effectiveness of test bench testing, playing a crucial supporting role in aero-engine performance testing, fault diagnosis, and optimized design. Currently, the mainstream technical solutions in pipeline flow characteristic prediction for environmental simulation systems can be categorized into three types. Each type has significant limitations and lacks a mechanism- and data-driven optimization system, making it difficult to meet the complex requirements of modern test bench testing.

[0003] Existing technologies are widely used in the field of aero-engine test bench research, but none have overcome the core limitation of "separation of mechanism and data," failing to achieve a balance between prediction accuracy, adaptability to operating conditions, and real-time computational efficiency. Existing technologies either suffer from insufficient accuracy due to ignoring actual disturbances, poor generalization due to a lack of mechanism support, or lack of real-time performance due to high computational demands, making it difficult to meet the precise control requirements of modern test benches for complex operating conditions. Therefore, a pipeline flow characteristic prediction technology solution that can overcome the above shortcomings is urgently needed. The shortcomings / deficiencies of existing technologies are as follows: Problem 1: Insufficient adaptability to transient states, making it difficult to adapt to conditions such as sudden valve opening and closing, and flow fluctuations. The core characteristic of transient states is "dynamic change," specifically manifested as sudden valve opening and closing (e.g., the opening of an 814 disc valve increases from 0° to 18° within 1 second), rapid flow fluctuations (increasing from 80 kg / s to 220 kg / s in a short time), and a wide temperature range (-65℃ to 350℃). Existing purely theoretical methods do not take these dynamic factors into account, and single machine learning models have insufficient learning capacity for transient state samples, resulting in flow parameter predictions that cannot fully meet the requirements of engineering experiments. Problem 2: Weak anti-interference capability. The test bench environment contains various external interferences, including electromagnetic noise, airflow disturbances, and sensor errors, which can distort the original data. Existing methods do not have targeted anti-interference mechanisms, purely theoretical methods cannot correct data errors, and single machine learning models are easily misled by noisy data, further amplifying prediction bias and leading to poor reliability of experimental data. Problem 3: Difficulty in achieving synergistic optimization of accuracy and efficiency. Purely theoretical methods have limited accuracy, while CFD simulations and single machine learning methods are inefficient and cannot meet the needs of real-time testing. Accuracy and efficiency cannot be simultaneously achieved, presenting a clear "either / or" dilemma: purely theoretical methods have high computational efficiency (reaching the millisecond level) but insufficient accuracy; CFD simulations have high accuracy but extremely low efficiency (remaining offline); single machine learning methods have moderate accuracy but require significant preprocessing time (reaching the minute level) and have poor generalization ability. However, test bench testing requires both controlling prediction errors to within 8% and providing real-time results (achieving second-level response), a core requirement that existing methods cannot meet.

[0004] The aforementioned deficiencies directly lead to low reliability of environmental simulation test data, which in turn affects the performance evaluation, operating condition optimization, and fault diagnosis of aero-engines, and restricts the technological upgrade of test bench testing. Therefore, there is an urgent need for a prediction method that integrates mechanisms and data, balances accuracy and efficiency, and is adaptable to complex scenarios. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a method for predicting pipeline flow characteristics in an environmental simulation system, aiming to overcome the following three key technical challenges: 1. Adaptation challenges of transitional operating conditions: Solve the problem of large deviations between pipeline outlet pressure and temperature prediction caused by sudden valve opening and closing (opening and closing time 1~13s), rapid flow fluctuations (80~220kg / s), and large temperature changes (-65℃~350℃), and bridge the adaptation gap between theoretical models and actual transitional operating conditions.

[0006] 2. External interference suppression problem: Solve the problem of reduced prediction accuracy caused by external interference such as electromagnetic noise, airflow disturbance, and sensor error in the test environment, and enhance the model's fault tolerance and robustness to distorted data.

[0007] 3. Balancing the challenges of accuracy and efficiency: Resolving the contradictions between the poor accuracy of purely theoretical methods, the low efficiency of CFD simulation, and the poor generalization ability of single machine learning methods, achieving the dual goals of "second-level response + error ≤ 8%" to meet the requirements of real-time monitoring in test bench experiments.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: a method for predicting the flow characteristics of a pipeline network in an environmental simulation system, comprising the following steps: S1. Combining theoretical derivation and CFD numerical simulation, we explore the evolution mechanism of pressure drop, outlet pressure and outlet temperature of the environmental simulation system under transient conditions. S2. Based on the investigation in S1, the flow velocity time series data of the pipeline outlet was collected through the transition state test data, and the calculation formulas and steps for the pipeline outlet pressure and temperature of different pipeline structures were derived based on the equations of mass conservation, momentum conservation and energy conservation. The collected flow velocity time series data was preprocessed. S3. Perform feature filtering on the preprocessed flow velocity time series data, and obtain key features by calculating the correlation between all filtered features and the outlet temperature correction coefficient and the outlet pressure correction coefficient. S4. Based on the derived calculation formulas and steps for pipeline outlet pressure and temperature, corrections are made in conjunction with transition state test data. A correction coefficient prediction model for outlet pressure and outlet temperature is established through machine learning algorithms. Key features are used as the input to the correction coefficient prediction model, and outlet temperature correction coefficient and outlet pressure correction coefficient are used as the outputs of the correction coefficient prediction model. S5. Use the correction coefficient prediction model to predict the flow characteristics of the pipeline network.

[0009] The beneficial effects of this invention are: The core concept of this invention is to construct a hybrid prediction framework of "theoretical research + CFD numerical simulation + machine learning model correction": First, by combining theoretical derivation and CFD numerical simulation, the evolution mechanism of pressure drop, outlet pressure, and outlet temperature of the environmental simulation system under typical transient conditions is explored in depth, and the sensitivity of each influencing factor is systematically analyzed. On this basis, time-series data of pipeline outlet flow velocity and velocity gradient are obtained through experimental data collection. Based on the equations of mass conservation, momentum conservation, and energy conservation, theoretical calculation formulas and steps for outlet pressure and temperature of different pipeline structures (one inlet and one outlet, one inlet and two outlets, one inlet and four outlets, and exhaust diffuser) are derived. Subsequently, by combining transient experimental data and steady-state flow parameter characteristics, a correction coefficient prediction model for outlet pressure and temperature is established through machine learning algorithms, forming common parameter rules for system flow characteristics. Finally, a pipeline flow characteristic prediction method based on a black-box model offline algorithm is proposed to accurately compensate for the deviation in outlet parameter prediction caused by rapid flow changes and significant external noise interference in engine transient tests, achieving high-precision and high-stability flow characteristic prediction.

[0010] This invention specifically focuses on the testing scenarios of aero-engine test benches and is applicable to the complete piping system of the test bench (covering the intake piping network PB1 and PB2 subsystems, the PC mixer system, the PD exhaust and extraction system, and the core equipment exhaust diffuser). Its core application scenario is the transitional operating condition—that is, the complex test process involving sudden valve opening and closing, rapid flow fluctuations, and dynamic adjustments in temperature and pressure. It aims to achieve high-precision real-time prediction of key flow parameters such as piping outlet pressure and temperature, providing reliable data support for aero-engine performance testing, operating condition simulation, and fault early warning, ensuring the accuracy of test results and the scientific nature of engineering decisions. Attached Figure Description

[0011] Figure 1 This is a flowchart of the method of the present invention.

[0012] Figure 2 This is a schematic diagram of a straight pipe cavity in a single-inlet-southlet pipe cavity model.

[0013] Figure 3 This is a schematic diagram of a pipeline cavity model with a 90° elbow and a valve in a single-inlet and single-outlet pipeline cavity.

[0014] Figure 4 This is a schematic diagram of a pipeline cavity with one inlet and two outlets.

[0015] Figure 5 This is a schematic diagram of a pipeline cavity with one inlet and four outlets.

[0016] Figure 6 This is a schematic diagram of a two-inlet, one-outlet pipeline cavity. Detailed Implementation

[0017] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0018] Example like Figure 1 As shown, this invention provides a method for predicting the flow characteristics of a pipeline network in an environmental simulation system, the implementation of which is as follows: S1. Combining theoretical derivation and CFD numerical simulation, we explore the evolution mechanism of pressure drop, outlet pressure and outlet temperature of the environmental simulation system under transient conditions. S2. Based on the investigation in S1, the flow velocity time series data of the pipeline outlet was collected through the transition state test data, and the calculation formulas and steps for the pipeline outlet pressure and temperature of different pipeline structures were derived based on the equations of mass conservation, momentum conservation and energy conservation. The collected flow velocity time series data was preprocessed. S3. Feature screening is performed on the preprocessed flow velocity time series data. By calculating the correlation between all screened features and the outlet temperature correction coefficient and the outlet pressure correction coefficient, key features are obtained. The implementation method is as follows: Based on the preprocessed flow velocity time series data, key features related to the outlet temperature correction coefficient and the outlet pressure correction coefficient are screened. Combined with variable sensitivity analysis, a general feature system for cross-straight pipes, tees, two-stream flow, and exhaust diffusers is established to complete the unified feature screening across the pipeline network structure and obtain key features. S4. Based on the derived calculation formulas and steps for pipeline outlet pressure and temperature, corrections are made using transient state test data. A prediction model for the correction coefficients of outlet pressure and temperature is then established using a machine learning algorithm. Key features serve as the input to the prediction model, while the outlet temperature correction coefficient and outlet pressure correction coefficient serve as the output. For example, the key input feature is: inlet static temperature. T in Export speed Import speed Sudden change time, inlet static pressure P inKey features selection: Based on the results of the variable sensitivity analysis in the technical disclosure document, "inlet static temperature (relative importance 76.46%), outlet velocity (14.7%), inlet velocity (4.05%), abrupt change time (2.83%), and inlet static pressure (1.95%)" were selected as input features for the GBDT model; Output targets: outlet temperature correction coefficient and outlet pressure correction coefficient; Prediction process: Input the above five features for the new operating condition; Model output: outlet temperature correction coefficient and outlet pressure correction coefficient; Model output: "pipeline outlet pressure correction coefficient" and "pipeline outlet temperature correction coefficient" were used as model outputs; Equivalent parameter calculation: Substitute the correction coefficients into the theoretical formula for outlet pressure / temperature of the exhaust diffuser to obtain the equivalent outlet pressure and equivalent outlet temperature.

[0019] S5. Predicting pipeline flow characteristics using a correction coefficient prediction model, including: Using a multilayer perceptron (MLP) model, the first equivalent outlet pressure is obtained by predicting the outlet pressure of a typical pipeline network, in order to predict the flow characteristics of the pipeline network. By using the gradient boosting decision tree model GBDT for the exhaust diffuser cavity, the second equivalent outlet pressure and equivalent outlet temperature are obtained by predicting the outlet pressure / temperature of the exhaust diffuser, so as to predict the flow characteristics of the pipeline network.

[0020] In this embodiment, the calculation formula and steps for the pipeline outlet pressure of different pipeline network structures are derived in the multilayer perceptron (MLP) model, as follows: Input pipeline length Pipe inner diameter Local drag coefficient and pipeline inlet cross-sectional pressure ; Obtain time-series flow velocity data at the pipeline outlet from the transition state test data; calculate the time-varying Reynolds number based on the flow velocity time-series data. Based on time-varying Reynolds number Iterative solution of dynamic friction coefficient Based on dynamic friction coefficient Flow velocity time series data Pipe length Pipe inner diameter and local drag coefficient The total pressure difference of the pipeline under unsteady flow was calculated. Based on the pressure at the pipe inlet section Total pressure difference in unsteady flow of pipeline The pipeline outlet pressure was calculated. .

[0021] In this embodiment, a multilayer perceptron (MLP) model is used to predict the outlet pressure of a typical pipeline network and calculate the first equivalent outlet pressure. This includes the following steps: extracting key features from the flow velocity time-series data under the new operating conditions; inputting the extracted key features into the MLP model to predict the outlet pressure correction coefficient; and combining the derived calculation formulas and steps for pipeline outlet pressure and temperature with the outlet pressure correction coefficient to calculate the first equivalent outlet pressure. For the flow velocity time series data under the new operating conditions, the same feature engineering method (such as time series differencing) as in the training phase is used. The extracted key features are input into the trained MLP model to predict the "pipeline outlet pressure correction coefficient". Combining the derived theoretical formula for pipeline outlet pressure, the correction coefficient is substituted to calculate the corrected pipeline outlet pressure (i.e., the first equivalent outlet pressure). Training phase: Six key features, including "valve opening and closing time, valve flow rate, valve temperature, pipe diameter, valve opening and closing method, and pipe length", are used as inputs; the "outlet pressure correction coefficient" is used as the output. Prediction phase: Input the same features for the new operating conditions → the trained MLP model outputs the correction coefficient → substitutes it into the outlet pressure formula → obtains the "first equivalent outlet pressure".

[0022] In this embodiment, the calculation formula and steps for the pipeline outlet pressure of different pipeline network structures are derived in the GBDT decision tree model of the exhaust diffuser cavity, as follows: Input import speed Imported static temperature T in Imported static pressure P in And to acquire time-series data of pipeline outlet flow velocity from transition state test data acquisition pipeline. ; Calculate the total inlet temperature T 0out and total pressure P 0out Based on total imported temperature T 0out and total pressure P 0out The total outlet temperature was calculated. T 0out and total pressure P 0out ;Analyze and correct the efficiency of the exhaust diffuser and total pressure recovery coefficient η p Based on the analysis results and total export temperature T 0out and total pressure P 0out Calculate the outlet static temperature T out and imported static pressure Pout .

[0023] In this embodiment, the Gradient Enhancement Decision Tree (GBDT) model for the exhaust diffuser cavity, by predicting the exhaust diffuser outlet pressure / temperature, obtains the second equivalent outlet pressure and equivalent outlet temperature, including the following steps: Based on the derived calculation formulas and steps for pipeline outlet pressure under different pipeline network structures, and combined with transient state test data, corrections were made. Based on the correction results, the second equivalent outlet pressure was obtained using the gradient boosting decision tree model (GBDT) of the exhaust diffuser cavity. and equivalent outlet temperature .

[0024] The present invention will be further described below.

[0025] In this invention, the method for predicting pipeline flow characteristics in an environmental simulation system includes four core steps, which combine MATLAB / SIMULINK to construct a simulation environment. The process is as follows: I. Data Acquisition and Preprocessing: Constructing a Dynamic Operating Condition Database Data source: Collect flow rate time series data (core input) from engine transient test, and simultaneously record outlet pressure and temperature test values ​​(target output).

[0026] Preprocessing: Denoising the raw data (e.g., wavelet denoising, Kalman filtering) to eliminate external noise interference; enriching data features through dimensionality expansion (e.g., time series differencing, sliding window feature extraction) to provide a foundation for subsequent analysis.

[0027] II. Feature Engineering: Screening Key Feature Parameters Principal component analysis (PCA) was used to select features from the denoised data. Calculate the correlation between all features and the "outlet temperature correction coefficient and outlet pressure correction coefficient", and screen out the key features that are significantly correlated (such as flow rate change, flow fluctuation amplitude, environmental noise intensity, etc.) to reduce model complexity and improve prediction accuracy.

[0028] III. Model Building and Validation: Machine Learning-Driven Correction Coefficient Prediction Model selection: Machine learning algorithms (multilayer perceptron (MLP) neural network and BOA-GBDT model) are used to construct a prediction model with "key feature parameters" as input and "correction coefficients" as output.

[0029] Training and Validation: Divide the dataset into a training set (80%) and a test set (20%). Train the model on the training set and optimize the hyperparameters using grid search or Bayesian methods. Evaluate the model performance (such as mean squared error MSE and coefficient of determination R2) on the test set to ensure the model's generalization ability.

[0030] IV. High-precision outlet pressure / temperature prediction: Application of correction coefficients Using a trained correction coefficient model, feature extraction and prediction are performed on the flow velocity time series data under the new operating conditions to obtain the correction coefficient; then, combined with the mathematical expressions for outlet pressure and temperature, the corrected outlet pressure / temperature values ​​are calculated; finally, the model accuracy is verified by comparing and analyzing the results with experimental values.

[0031] In this embodiment, the theoretical model for the outlet pressure and outlet temperature of the pipeline cavity is derived as follows.

[0032] 1. One-inlet-one-outlet pipe cavity model The inlet and outlet pipe cavity model mainly includes a straight pipe cavity and a pipe cavity with fittings (90° elbows, valves, etc.), as shown in the schematic diagram. Figure 2 and Figure 3 As shown.

[0033] Gas law: (1) Mass conservation (continuity) equation: (2) Energy conservation equation: (3) in, The pressure of the gas inside the pipe is expressed in Pa. Indicates the temperature of the gas inside the pipe, in K; This indicates the mass flow rate of the gas inside the pipe, in kg / s. The volume of the pipe is expressed in meters (m). 3 ; q mino This indicates the outlet flow rate of the pipeline, in kg / s; q mout This indicates the outlet flow rate of the pipeline, in kg / s; Instantaneous velocity, expressed in m / s; This represents the specific heat capacity of a gas at constant pressure, in J / (kg·K). This indicates the density of a gas, expressed in kg / m³. 3 ; R The gas constant representing the air inside the pipe; Indicates the inner diameter of the pipe, in meters (m). The thermal conductivity of a gas is expressed in W / (m²). 2 .K); The convective heat transfer coefficient of a gas is expressed in W / (m³). 2 .K); The temperature of the channel wall is expressed in Kelvin (K).

[0034] Total pressure difference in unsteady flow within the pipeline: (4) The first term in equation (6) is the resistance loss term, which includes frictional resistance loss and local resistance loss, and the second term is the inertia term.

[0035] Pipeline outlet pressure: (5) Among them, dynamic friction resistance coefficient Ignoring the influence of pipe wall roughness and assuming it is a smooth pipe, the following two formulas are used for dynamic calculation to obtain the result.

[0036] When the Reynolds number is The calculation is performed using the Blasius formula, namely: (6) When the Reynolds number is The calculation is performed using the Nikolaiz formula, namely: (7) Among them, the time-varying Reynolds number It is related to instantaneous velocity, that is: (8) in, This represents the total pressure difference in the pipeline during unsteady flow, expressed in Pa. This indicates the pressure at the pipe inlet cross-section, in Pa; This indicates the pressure at the pipe outlet section, in Pa. This represents the time-varying frictional resistance coefficient, which is related to the Reynolds number and the pipe wall roughness. This represents the local resistance coefficient, which is related to abrupt changes in pipe shape. The length of the pipe is expressed in meters (m). This indicates the absolute roughness of the pipe wall, in mm; Represents the Reynolds number, which is dimensionless.

[0037] Total temperature difference in unsteady flow of the pipeline: (9) Among them, the convective heat transfer coefficient It is obtained through dynamic calculation using the Dittus-Boelter formula, namely: (10) in, The thermal conductivity of a gas is expressed in W / (m²). 2K), Re represents the time-varying Reynolds number (dimensionless), used to determine the flow state (turbulent / laminar), Pr represents the Prandtl number (dimensionless). This represents the heat flux density of the pipe wall, in W / m³. 2 .

[0038] Pipeline outlet temperature: (11) in, The total temperature difference in the pipeline during unsteady flow is expressed in K. This indicates the temperature at the pipe inlet section, in K. This indicates the temperature at the pipe outlet section, in K.

[0039] 2. One-inlet, two-outlet pipe cavity model The one-inlet, two-outlet pipe cavity model mainly includes a three-way cavity, such as... Figure 4 As shown. Its gas law is the same as that of a gas with one inlet and one outlet, so it will not be described here. However, the mass conservation equation and the energy conservation equation are different, and will be derived in detail below.

[0040] Mass conservation (continuity) equation: (12) The temperature distribution in unsteady flow is derived based on the energy conservation equation and the transient heat transfer equation, namely: (13) in, This represents the Dirac function and indicates the location of the heat transfer point in the branch pipe.

[0041] Total pressure difference in unsteady flow within the pipeline: (14) in, Indicates the equivalent length of the branch pipe, in meters (m). i This represents the parameters corresponding to the pipe inlet (in) and the two outlets (out1, out2). The outlet pressures (out1, out2) correspond to the pipe's outlet pressures. (15) (16) The outlet temperatures of the pipes corresponding to the two outlets, out1 and out2, are as follows: (17) (18) in, and These represent the pipeline outlets. and pipeline outlet The total pressure difference in unsteady flow, and These represent the pipeline outlets. and pipeline outlet The time-varying frictional resistance coefficient, and These represent the pipeline outlets. and pipeline outlet The corresponding pipe length, and These represent the pipeline outlets. and pipeline outlet The inner diameter of the pipe, and These represent the pipeline outlets. and pipeline outlet The local drag coefficient, and These represent the pipeline outlets. and pipeline outlet instantaneous speed, and These represent the pipeline outlets. and pipeline outlet The outlet temperature, and Indicates pipeline outlet and pipeline outlet The gas convection heat transfer coefficient.

[0042] 3. One-inlet, four-outlet pipe cavity model A one-inlet, four-outlet pipe cavity model mainly includes two flow cavities, such as... Figure 5 As shown. Its gas law is the same as that of a gas with one inlet and one outlet, so it will not be described here. However, the mass conservation equation and the energy conservation equation are different, and will be derived in detail below.

[0043] Mass conservation (continuity) equation: (19) The temperature distribution in unsteady flow is derived based on the energy conservation equation and the transient heat transfer equation, i.e. (20) in, Represents the Lark function, used to locate the heat transfer position of the branch pipe; i This represents the parameters corresponding to the pipeline inlet (in) and the four outlets (out1, out2, out3, out4).

[0044] Total pressure difference in unsteady flow within the pipeline: (twenty one) Pipeline outlet pressure: (twenty two) Pipeline outlet temperature: (twenty three) in, Indicates pipeline The total pressure difference in unsteady flow, Indicates pipeline The time-varying frictional resistance coefficient, Indicates pipeline Length, Indicates pipeline inner diameter, Indicates pipeline The local drag coefficient, Indicates pipeline instantaneous speed, Indicates pipeline The outlet temperature, Indicates the convective heat transfer coefficient at different pipe outlets , i Indicates pipeline inlet in and four exits out 1. out 2. out 3 and out The number 4.

[0045] 4. Two-inlet, one-outlet pipe cavity model The two-inlet, one-outlet pipe cavity model mainly includes the C1 mixer cavity, such as... Figure 6 As shown.

[0046] Mass conservation (continuity) equation: (twenty four) The temperature distribution in unsteady flow is derived based on the energy conservation equation and the transient heat transfer equation, namely: (25) Total pressure difference in unsteady flow within the pipeline: (26) in, i This indicates the parameters corresponding to the pipeline inlet (in1), in2, and outlet (out).

[0047] Outlet pressure of the outlet pipeline: (27) in, This represents the equivalent pipe length from the first inlet to the outlet, in meters (m).

[0048] Pipeline outlet temperature: (28) in, Indicates pipeline inlet The cross-sectional pressure, Indicates pipeline outlet The total pressure difference in unsteady flow, Indicates pipeline outlet The time-varying frictional resistance coefficient, Indicates pipeline outlet The length of the pipe, Indicates pipeline outlet The inner diameter of the pipe, Indicates pipeline outlet The time-varying frictional resistance coefficient, Indicates pipeline outlet instantaneous speed, This indicates the flow rate at the inlet of the first pipe. This indicates the inlet temperature of the first pipe. This indicates the flow rate at the inlet of the second pipe. This indicates the inlet temperature of the first pipe. k Indicates the first k One pipeline inlet, Indicates the outlet convective heat transfer coefficient. This indicates the average temperature of the inlet pipe.

[0049] In this embodiment, the outlet pressure calculation steps in the multilayer perceptron neural network model (i.e., the multilayer perceptron MLP model) of the pipeline cavity are as follows: The model parameter calibration and calculation steps for the outlet pressure calculation models of the one-in-one-out, one-in-two-out, one-in-four-out, and two-in-one-out pipelines in the test bench system are as follows: Figure 5 As shown in the figure, the model parameter calibration and calculation steps are as follows: Input conditions: pipe length L, pipe diameter d, local resistance coefficient Import pressure P in and temperature T in， and flow velocity time series data ; Calculate the time-varying Reynolds number ; Dynamic friction coefficient ; Total pressure difference in unsteady flow of pipeline ; according to Calculate the pressure at the pipe outlet section. .

[0050] In this embodiment, the valve flow rate, valve opening and closing time, pipe diameter, pipe length, valve temperature, and valve opening and closing method in the exhaust system of a certain test stand for the 2024-2025 aero-engine test simulation test in the input model are shown in Table 1. Table 1 is a descriptive statistical table of each variable.

[0051] Table 1

[0052] Table 1 shows that the valve flow rate ranges from 80 to 220 kg / s, the valve opening and closing time ranges from 1 to 13 s, the pipe diameter ranges from 1 to 3 m, the pipe length ranges from 0.5 to 26 m, the valve temperature ranges from 10 to 180℃, and the valve opening and closing modes include pulsating, linear, and quadratic changes. An MLP neural network model was used to analyze the relative importance of each input variable, and the results are as follows: Figure 6 As shown, valve flow rate has the highest relative importance to the model, at 0.39, followed by valve opening and closing time at 0.31. The sensitivity of the six influencing factors, in descending order, is: valve flow rate > valve opening and closing time > pipe diameter > valve opening and closing method > pipe length > valve temperature.

[0053] In this embodiment, the Multilayer Perceptron (MLP) is a feedforward neural network model based on error backpropagation. It achieves a complex mapping from input to output through a hierarchical nonlinear transformation structure. In time series data prediction tasks, MLP demonstrates powerful pattern recognition capabilities. Its core mechanism lies in constructing a high-level abstract representation of features through forward propagation and globally optimizing the network weights using the backpropagation algorithm, ultimately achieving an accurate fit to the prediction function. Compared to traditional time series models, the core advantage of MLP is its ability to construct nonlinear functions of arbitrary complexity based on the universal approximation theorem. It does not require pre-defining the specific form of time-series dependencies. Through the distributed representation of hidden layers and the nonlinear transformation of activation functions, it automatically captures the deep spatiotemporal dynamics inherent in the data, ultimately forming a highly generalizable universal function approximator.

[0054] Due to factors such as rapid changes in flow rate in the test chamber and significant interference from external noise during engine transient testing, deviations in the outlet pressure and temperature transmitted by the pipeline model are required. Based on theoretically derived formulas and calculation steps for pipeline outlet pressure and temperature, and combined with experimental data, corrections are made to obtain a prediction model with correction coefficients for outlet pressure and temperature. Then, the first equivalent outlet pressure and equivalent outlet temperature are calculated. (29) The pipeline outlet pressure correction factor is the valve opening and closing time ( ), valve flow rate ( ), valve temperature ( ), pipe diameter ( ), Valve opening and closing method ( ), pipe chief ( The six independent variables, ranked in order of sensitivity, are: valve flow rate > valve opening / closing time > pipe diameter > valve opening / closing method > pipe length > valve temperature. Using the pipeline outlet pressure correction coefficient as the dependent variable, a 10-fold cross-validation method is employed for training and evaluation, ultimately constructing the MLP model as follows: (30) in, This indicates the first equivalent export pressure. This represents the export pressure correction factor. Indicates the pipeline outlet pressure. This represents a multilayer perceptron. Indicates the valve opening and closing time. Indicates valve flow rate. Indicates valve temperature. Indicates pipe diameter. Indicates the valve opening and closing mode. Indicates the length of the pipe.

[0055] The prediction of correction coefficients needs to be completed in three stages: data preprocessing, feature engineering, and model building and validation. The core issue is to solve the problem of outlet pressure deviation caused by "rapid flow changes and external noise interference" in engine transient tests. The detailed process is as follows: 1. Data Acquisition and Preprocessing: Constructing a Dynamic Operating Condition Database Data Source: Flow velocity time-series data (core input) collected from engine transient state tests, with outlet pressure recorded simultaneously. Preprocessing: Wavelet denoising was performed on the raw data to eliminate external noise interference; time series differencing was used to expand the dimensionality of the data and enrich its features, providing a foundation for subsequent analysis.

[0056] 2. Feature Engineering: Screening Key Feature Parameters Principal component analysis (PCA) was used to screen features in the denoised data: the correlation between all features and the "exit pressure correction coefficient" was calculated, and key features with significant correlation (flow rate change) were screened out to reduce model complexity and improve prediction accuracy.

[0057] 3. Model Building and Validation: Machine Learning-Driven Correction Coefficient Prediction Model Selection: An MLP neural network model is adopted, using "key feature parameters" as input and "correction coefficients" as output to construct a prediction model. Training and Validation: The dataset is divided into a training set (80%) and a test set (20%). The model is trained on the training set using 10-fold cross-validation; the model performance is evaluated on the test set (mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2This ensures the model's generalization ability.

[0058] 4. High-precision export pressure prediction: application of correction coefficients Feature extraction and prediction were performed on the flow velocity time series data under the new operating conditions to obtain the pipeline outlet velocity. Then, combined with the mathematical expression for outlet pressure and a trained correction coefficient model, the corrected outlet pressure was calculated. Finally, the model accuracy was verified by comparing and analyzing the results with experimental values. To evaluate the model's generalization ability and stability, a 10-fold cross-validation method was used to systematically validate the MLP pressure correction coefficient prediction model. Cross-validation is an important model evaluation technique; by repeatedly training and validating on multiple subsets of data, it can more reliably estimate the model's performance on unknown data.

[0059] In terms of data partitioning strategy, the total number of data points were divided into training and test sets in an 8:2 ratio. A 10-fold cross-validation method was used, randomly dividing the training set into 752 equal-sized subsets. In each cross-validation iteration, 602 subsets were used as training data, and the remaining 150 subsets were used as validation data. This process was repeated 10 times to ensure that each subset had at least one opportunity to serve as the validation set. Finally, final validation was performed on the entire dataset.

[0060] In this embodiment, a trained neural network model is used to predict a set of unknown operating conditions that were not included in the training: valve opening and closing time (7s), valve flow rate (220kg / s), valve temperature (323.15K), pipe diameter (1.8m), valve opening and closing mode (pulsating change), and pipe length (26m). Correction coefficients are then used to predict these parameters. Combined with the mathematical expression for the outlet pressure, the corrected outlet pressure is compared with the original outlet pressure and the actual outlet pressure. The corrected outlet pressure shows a strong positive correlation with the experimental value, and the overall trend is close to the ideal state of linear fitting (the slope of the trend line is close to 1). The maximum relative error between the calculated and experimental values ​​of the pipeline outlet pressure before correction was 35%, while after correction, the maximum relative error was 2.9%, a decrease of 32.1%. This meets the requirement that the relative error of the predicted flow parameters does not exceed 8%.

[0061] In this embodiment, the theoretical model for the outlet pressure and outlet temperature of the exhaust diffuser is derived.

[0062] The exhaust system is simulated during aircraft engine test runs, and the inlet static pressure is usually known. T in Imported static pressure P in Obtain the static temperature at the outlet of the exhaust diffuser. T out and outlet static pressure P out The calculation steps are as follows: Step 1: Based on the imported static temperature T in Imported static pressure P in Calculate the total inlet temperature T 0out and total pressure P 0out The calculation formulas are as shown in equation (36) and equation (37).

[0063] Combining the laws of mass conservation and energy conservation, we derive expressions for the outlet pressure and temperature of the exhaust diffuser; that is, the mass conservation equation is the foundation for calculating the outlet parameters of the exhaust diffuser. For steady flow, mass conservation can be expressed as: (31) in, ρ This indicates the density of a gas, expressed in kg / m³. 3 ; A Represents the cross-sectional area, m 2 ; v Indicates velocity, m / s.

[0064] The energy conservation equation is used to calculate the outlet temperature. For an ideal gas, neglecting heat exchange and irreversible losses, the total enthalpy is conserved: (32) in, h Specific enthalpy, J / kg; For ideal gases such as air, specific enthalpy can be expressed as: (33) in, C p This represents the specific heat capacity at constant pressure, in J / (kg·K); T Indicates temperature; K.

[0065] Import total temperature: (34) in, T 0i Indicates the total import temperature, in K; T i Indicates imported static temperature, of which i For example, in represents the engine exhaust nozzle, in1 and in2 are the two inlets for the two streams, and K represents the inlet location. v i Indicates the inlet velocity, where, for example, in is the engine exhaust nozzle, in1 and in2 are the two inlets of the two streams, in m / s; C p This represents the specific heat capacity at constant pressure, in J / (kg·K).

[0066] Total inlet pressure: (35) in, P 0i Indicates the total inlet pressure, in Pa; P i Indicates the inlet static pressure, in Pa; k This indicates the adiabatic index.

[0067] Step 2: Based on the total imported temperature T 0in and total pressure P 0in Calculate the total outlet temperature T 0out and total pressure P 0out The calculation formulas are shown in equations (38) and (39). During the entire gas flow process in the exhaust diffuser, the diffuser efficiency... η and total pressure recovery coefficient η p The main influences are related to the compressibility and viscosity of the airflow; that is, when Ma > 0.3, the compressibility effect needs to be considered, as well as the low... R At time e, viscous forces dominate, and boundary layer thickening leads to increased losses. This invention uses Fluent simulation analysis to assess diffuser efficiency. η Total pressure recovery coefficient η p Make corrections.

[0068] (36) (37) Total temperature at the exit: (38) in, T 0out Indicates the total outlet temperature, in K; v 2 indicates the exit velocity, in m / s; η This indicates the diffuser efficiency.

[0069] Total pressure at the outlet: (39) in, P 0out Indicates the total outlet pressure, in Pa; η p This represents the total pressure recovery coefficient.

[0070] Step 3: Based on the total outlet temperature T 0out and total pressure P 0out Calculate the outlet static temperature Tout Imported static pressure P out The calculation formulas are as shown in equation (40) and equation (41).

[0071] Outlet static pressure: (40) in, T out Indicates the outlet static temperature, K.

[0072] Static temperature at the outlet: (41) in, P out Indicates the outlet static pressure, in Pa. Indicates the export speed. Indicates the total import temperature. This represents the specific heat capacity of a gas at constant pressure. Indicates the import speed. This indicates the efficiency of the exhaust diffuser. k The adiabatic index indicates the thermal properties of a gas. η p This represents the total pressure recovery coefficient.

[0073] In this embodiment, the calculation steps for outlet pressure and temperature in the gradient boosting decision tree model (GBDT) of the exhaust diffuser cavity are as follows: The derivation model reveals that when the airflow properties are constant, the outlet pressure of the transition pipe is mainly affected by the airflow velocity distribution within the exhaust diffuser, as detailed in equations (40) and (41). The model parameter calibration and calculation steps for the above exhaust diffuser outlet pressure and temperature calculation model are as follows: 1. Input conditions: Import speed u 1. Imported static temperature T 1. Imported static pressure P 1 and flow velocity time series data ; 2. Calculate the total inlet temperature T 0in and total pressure P 0in : ; 3. Calculate the total outlet temperature T0out and total pressure P0out: ; ; 4. Correct diffuser efficiency through Fluent simulation analysis. η Total pressure recovery coefficient η p ; 5. Calculate the outlet static temperature Tout Imported static pressure P out : ; .

[0074] In this embodiment, a total of 10,810 valid data points were input into the exhaust system of an aero-engine test simulation test on a certain test stand. These data included the inlet static pressure, inlet static temperature, inlet velocity, outlet velocity, and abrupt change time of the exhaust diffuser. See Table 3 for details. Table 2 is a descriptive statistical table of each variable.

[0075] Table 2

[0076] Gradient boosting tree (GBDT) was used to conduct a relative importance analysis of each input variable. The analysis showed that inlet static temperature had the highest relative importance (76.46%), indicating that it is the most critical factor affecting the pressure correction coefficient. The outlet velocity was the second most important, with a relative importance of 14.7%. The remaining three factors had importance levels below 5%, with inlet velocity at 4.05%, abrupt change time at 2.83%, and inlet static pressure at 1.95%. The order of sensitivity of the five factors to the pressure correction coefficient was: inlet static temperature > outlet velocity > inlet velocity > abrupt change time > inlet static pressure. Abrupt change time had the highest relative importance (39.97%), indicating that time is the decisive variable affecting the temperature correction coefficient. Inlet static temperature and outlet velocity had relative importance levels of 29.39% and 15.24%, respectively, ranking second and third. Inlet velocity and inlet static pressure had lower importance levels, at 11.96% and 3.44%, respectively. The order of sensitivity of the five influencing factors to the temperature correction coefficient is: abrupt change time > inlet static temperature > outlet velocity > inlet velocity > inlet static pressure.

[0077] In this embodiment, Gradient Boosting Decision Tree (GBDT) is a forward stepwise additive model based on ensemble learning. It constructs a strong predictor by iteratively combining multiple weak learners (decision trees). The core mechanism of this method lies in gradually correcting the prediction residuals of the preceding models through a gradient descent strategy, ultimately achieving a continuous improvement in prediction accuracy. Compared to traditional regression models, GBDT's advantage lies in its ability to automatically learn the complex nonlinear mapping relationship between input and output, without requiring strict assumptions about the data distribution. By gradually correcting prediction errors through a forward stepwise algorithm and a gradient descent strategy, it ultimately constructs a highly adaptive strong predictor.

[0078] Gradient Boosting Decision Tree (GBDT) does not require the data to satisfy the stationarity assumption and can directly handle non-stationary industrial process data. The model uses an additive model representation. (42) in, Let represent the ensemble model after the m-th iteration, and v represent the learning rate, which controls the contribution of each tree. This represents the number of partitions in the m-th tree. This represents the output value of the corresponding leaf node. This indicates the region to be divided in the feature space.

[0079] Each newly added decision tree in each round specifically learns the negative gradient direction of the residuals of the preceding model. Through this round-by-round optimization mechanism, the model can effectively capture the complex nonlinear relationships and multi-factor interactions between features and the target variable. During model training, by limiting the depth of each tree and the number of leaf nodes, and introducing regularization techniques such as feature sampling and learning rate decay, overfitting is effectively avoided, ensuring the model's generalization ability on unknown data. This method does not require the data to meet specific statistical distribution assumptions and can directly handle the nonlinear and non-normal distribution characteristics commonly found in industrial data, providing a reliable technical means for high-precision prediction of exhaust system correction coefficients.

[0080] Due to factors such as rapid changes in flow rate in the test chamber and significant external noise interference during engine transient testing, deviations in the outlet pressure and temperature transmitted by the pipeline model are required. Based on theoretically derived formulas and calculation steps for pipeline outlet pressure and temperature, and combined with experimental data, corrections are made to obtain a prediction model with correction coefficients for outlet pressure and temperature. Then, the equivalent outlet pressure and equivalent outlet temperature are calculated, i.e.: (43) (44) in, This represents the equivalent export pressure, expressed in kPa. Represents the equivalent outlet temperature, in K; This represents the correction factor for the pipeline outlet pressure. This represents the correction factor for the outlet temperature of the pipeline network.

[0081] Based on the above explanation, it can be seen that the influencing factors of the outlet pressure correction coefficient and outlet temperature correction coefficient of the exhaust diffuser are the inlet static temperature (…). ), export speed ( ), import speed ( ), mutation time ( ), imported static pressure ( Five influencing factors were identified, with the sensitivity ranking as follows: inlet static temperature > outlet velocity > inlet velocity > abrupt change time > inlet static pressure, and abrupt change time > inlet static temperature > outlet velocity > inlet velocity > inlet static pressure. Using the pipeline outlet pressure correction coefficient and pipeline outlet temperature correction coefficient as dependent variables, a 10-fold cross-validation method was employed for training and evaluation. The final GBDT model was constructed as follows: (45) (46) in, This represents the correction factor for the pipeline outlet pressure. Indicates imported static temperature. Indicates the import speed. Indicates the export speed. Indicates the time of mutation. Indicates the imported static pressure. This represents the correction factor for the outlet temperature of the pipeline network.

[0082] The prediction of correction coefficients needs to be completed in three stages: data preprocessing, feature engineering, and model building and validation. The core issue is to solve the problem of outlet pressure / temperature deviation caused by "rapid flow changes and external noise interference" in engine transient tests. The detailed process is as follows: I. Data Acquisition and Preprocessing: Constructing a Dynamic Operating Condition Database Data Source: Time-series data of flow rate / volume rate from engine transient tests (core input), and simultaneous recording of outlet pressure and temperature test values ​​(target output). Preprocessing: Noise reduction (wavelet denoising) is performed on the raw data to eliminate external noise interference; data features are enriched through dimensionality expansion (time series differencing) to provide a foundation for subsequent analysis.

[0083] II. Feature Engineering: Screening Key Feature Parameters Principal component analysis (PCA) was used to screen features of the denoised data: the correlation between all features and the "outlet temperature correction coefficient and outlet pressure correction coefficient" was calculated, and key features with significant correlation (such as flow rate change rate, flow fluctuation amplitude, environmental noise intensity, etc.) were screened out to reduce model complexity and improve prediction accuracy.

[0084] III. Model Building and Validation: Machine Learning-Driven Correction Coefficient Prediction Model Selection: A machine learning algorithm (Gradient Boosting Tree (GBDT)) is used, with "key feature parameters" as input and "correction coefficients" as output, to construct a prediction model. Training and Validation: The dataset is divided into a training set (80%) and a test set (20%). The model is trained on the training set, and hyperparameters are optimized using grid search or Bayesian methods. The model performance (mean squared error MSE, coefficient of determination R) is evaluated on the test set. 2 This ensures the model's generalization ability.

[0085] IV. High-precision outlet pressure / temperature prediction: Application of correction coefficients Using a trained correction coefficient model, feature extraction and prediction are performed on the velocity / flow rate time series data under new operating conditions to obtain correction coefficients; then, combined with the mathematical expressions for outlet pressure and temperature, the corrected outlet pressure / temperature values ​​are calculated; finally, the model accuracy is verified by comparing and analyzing the results with experimental values.

[0086] This embodiment employs a grid search-based hyperparameter optimization method to tune the Gradient Boosting Tree (GBDT) model. This method systematically traverses the preset parameter space and, combined with 3-fold cross-validation and the principle of minimizing mean squared error (MSE), evaluates the performance of each parameter combination. This ensures comprehensive search coverage while avoiding local optima, thus guaranteeing the model's good generalization ability. The optimization process focuses on five key parameters: the number of decision trees (n_estimators), the learning rate (learning_rate), the minimum number of leaf samples (min_samples_leaf), the maximum number of splits (max_splits), and the feature sampling ratio (max_features). Specifically, the number of decision trees is set to 50-250 to balance model complexity and computational efficiency, the learning rate is set to 0.01-0.20 to coordinate model convergence speed and generalization performance, the minimum number of leaf samples is set to 1-15 to suppress overfitting, the maximum number of splits is set to 20, 50, and 100 respectively to control the depth of a single tree, and the feature sampling ratio is set to 0.6, 0.8, and 1.0 to improve model diversity and robustness. Optimization results show that the optimal parameter combination for the pressure correction coefficient model is n_estimators=250, learning_rate=0.1, min_samples_leaf=1, max_splits=100, max_features=0.6, corresponding to performance indices R²=0.9740 and RMSE=0.00143; the optimal parameter combination for the temperature correction coefficient model is n_estimators=250, learning_rate=0.15, min_samples_leaf=3, max_splits=100, max_features=0.6, corresponding to performance indices R²=0.9906 and RMSE=0.00185. The results indicate that both models exhibit optimal performance when the number of decision trees is relatively large (150–250) and the learning rate is moderate (0.1–0.15), effectively balancing accuracy and generalization ability, thus verifying the rationality and engineering applicability of this parameter optimization strategy.

[0087] The optimization results of both models jointly demonstrate that the ranking criterion based on cross-validation scores (mean squared error) can systematically select the optimal parameter combination, and the parameter selection range exhibits a regular characteristic of a relatively large number of decision trees, a moderate learning rate, and a fixed feature sampling ratio. This not only verifies the practicality of the loss minimization principle but also provides an efficient and stable model configuration scheme for engineering applications.

[0088] In this embodiment, the unknown operating condition prediction takes the 5-second transition from the engine's intermediate state to a low afterburner state as an example. The input boundary conditions are: change time (0-5s), inlet static pressure (69.44kPa-70.27kPa), inlet static temperature (628.6K-1216.63K), inlet velocity (129.07m / s-249.58m / s), and outlet velocity (70.26m / s-147.02m / s). Then, a trained prediction model is called to obtain the corrected outlet pressure and temperature. Before correction, the maximum relative error between the calculated and experimental values ​​of the pipeline outlet pressure was 3.7%, and after correction, the maximum relative error was 2.5%, a decrease of 30%. This meets the requirement that the relative error of the predicted flow parameters does not exceed 8%. Before correction, the maximum relative error between the calculated and experimental values ​​of the pipeline outlet temperature was 10.5%, and after correction, the maximum relative error was 5.4%, a decrease of 50%.

[0089] This invention overcomes existing technical bottlenecks in five dimensions—accuracy, adaptability, anti-interference, generalization ability, and efficiency—through a hybrid framework of "theoretical modeling + machine learning correction" and targeted technical design. Each advantage relies on the core technical solution to produce clear technical effects, as detailed below: The prediction accuracy is significantly improved, fully meeting engineering requirements: The core of this invention lies in accurately compensating for theoretical calculation deviations through machine learning models, forming a dual guarantee of "mechanism as a safety net + data correction". For ordinary pipeline networks, an MLP neural network is used to construct a pressure correction coefficient model, combined with 10-fold cross-validation optimization, reducing the maximum relative error of outlet pressure from 35% in traditional theoretical calculations to 2.9%, a reduction of 32.1%. For exhaust diffusers, through GBDT models and hyperparameter optimization (the optimal combination is n_estimators=250, learning_rate=0.15, etc.), the maximum relative error of outlet pressure is reduced from 5% to 2%, and the maximum relative error of temperature is reduced from 18% to 1.5%, reductions of 60% and 92% respectively. The prediction errors of all core parameters are strictly controlled within the technical target of 8%, far superior to the accuracy levels of pure theoretical methods (significant deviations), single machine learning methods (medium accuracy), and CFD simulations (offline high accuracy but not applicable in real time), providing reliable data support for aero-engine testing. With strong adaptability to transition states, this invention is well-suited for complex dynamic test scenarios: Existing technologies struggle to handle transition state conditions such as sudden valve opening and closing (1~13s), flow fluctuations (80~220kg / s), and dramatic temperature changes (-65℃~350℃). This invention enhances adaptability through two major technological designs: First, in the theoretical modeling stage, formulas for total pressure difference and total temperature difference in unsteady flow are derived, incorporating dynamic parameters such as time-varying Reynolds number and dynamic friction coefficient (Brashuus / Nikolaiz formula switching) to accurately characterize the transition state flow mechanism. Second, through variable sensitivity analysis, core influencing factors such as valve flow rate (relative importance 0.39), valve opening and closing time (0.31), and sudden change time of the exhaust diffuser (temperature correction coefficient importance 39.97%) are identified, enabling the MLP / GBDT model to specifically capture dynamic patterns and adapt to various transition state scenarios without additional adjustments, thus solving the problem of "good steady-state adaptability but large dynamic deviation" in traditional methods. With outstanding anti-interference capabilities, this invention ensures prediction stability under complex environments: Electromagnetic noise, airflow disturbances, and sensor errors in test bench experiments can easily lead to data distortion. This invention constructs a full-process anti-interference system: In the data preprocessing stage, wavelet denoising and Kalman filtering techniques are used to eliminate external noise, and time series differencing and sliding window feature extraction are used to optimize data quality. At the model level, the MLP / GBDT model is trained with 10-fold cross-validation and has strong noise tolerance, which can effectively offset the impact of distorted data on prediction results. From the validation results, the corrected model still maintains an R² value of over 0.97 (pressure correction) and over 0.99 (temperature correction) under noisy data input, which is significantly better than pure theoretical methods (which cannot correct errors) and single machine learning models (which are easily misled by noise), ensuring the stability of prediction function under complex working conditions.

[0090] With broad generalization capabilities, this invention covers unified modeling of multiple pipe network structures: Existing technologies often model specific pipe network structures individually, limiting their adaptability. This invention achieves multi-structure coverage through a unified theoretical framework and feature selection method: Theoretically, it derives the general conservation equations and outlet parameter calculation formulas for single-inlet-single-outlet, single-inlet-two-outlet, single-inlet-four-outlet, two-inlet-single-outlet, and exhaust diffusers, clarifying key parameters such as equivalent length and branch pipe heat transfer location for different structures; In the feature engineering stage, it uses PCA principal component analysis combined with variable sensitivity ranking to select key features common to multiple structures (such as flow rate change rate and inlet static temperature), enabling the hybrid model to achieve high-precision prediction without separate optimization for straight pipes, tees, two-stream structures, etc., significantly reducing modeling costs in engineering applications. Accuracy and efficiency are synergistically optimized to meet real-time testing requirements: This invention resolves the inherent trade-off between accuracy and efficiency in existing technologies. The theoretical calculation module rapidly outputs basic parameters based on three conservation equations, avoiding the offline computation costs of CFD simulations that can take hours or days. The machine learning correction module compensates only for deviations, and after optimization, the single-condition prediction response time is controlled within 5 seconds, achieving a computational efficiency more than 10 times higher than CFD simulation, while maintaining the high efficiency of millisecond-level theoretical calculations and the high accuracy of machine learning. Compared to purely theoretical methods (millisecond-level but low accuracy) and single machine learning methods (minute-level preprocessing and poor generalization), this invention achieves the dual goals of "second-level response + error ≤ 8%", fully adapting to the engineering requirements of real-time monitoring and immediate decision-making in test bench testing.

[0091] This invention forms a multi-level protection system based on a "hybrid prediction framework + targeted technical design," with core protection points focusing on innovative breakthroughs and auxiliary protection points ensuring the feasibility of technology implementation. The primary innovation lies in the hybrid prediction framework combining theoretical calculations and machine learning corrections. Scope of Protection: Based on the three conservation equations of fluid mechanics (mass, momentum, and energy), a theoretical model of unsteady pipeline flow is constructed. This model combines an MLP neural network and a GBDT model to compensate for deviations in the outlet pressure and temperature of the pipeline network and the exhaust diffuser, forming a collaborative architecture of "mechanism modeling for bottom-line adaptability + data-driven correction for improved accuracy." This framework fundamentally solves the core problem of the disconnect between mechanism and data in existing technologies, achieving a unity of accuracy, efficiency, and adaptability, and is the core support for all the technical effects of this invention.

[0092] Key optimization points: Unified feature selection and sensitivity adaptation method for cross-pipeline structures Scope of Protection: Based on principal component analysis (PCA) to screen key features significantly correlated with the outlet parameter correction coefficients, and combined with variable sensitivity analysis (such as ranking by valve flow rate > valve opening / closing time, inlet static temperature > outlet velocity, etc.), a general feature system is established across straight pipes, tees, two-stream systems, and exhaust diffusers. This improves the model's specificity and generalization ability, eliminating the need for separate modeling for different pipe network structures. This method solves the problem of limited adaptability in existing models and expands the application scenarios of the technology. Model optimization points: Targeted application of dual machine learning models and hyperparameter optimization strategies Scope of Protection: A targeted approach is adopted, using an MLP neural network to predict outlet pressure in ordinary pipeline networks and a GBDT model to predict pressure / temperature in exhaust diffusers. Hyperparameters (number of decision trees, learning rate, minimum number of leaf node samples, etc.) are optimized through grid search and cross-validation (10-fold / 3-fold) to determine the optimal parameter combination for both models, ensuring both accuracy and generalization ability. This design achieves precise matching of the machine learning model with different prediction scenarios, representing a key technological breakthrough in accuracy. Data protection points: Multi-stage data preprocessing and anti-interference process The protection scope includes wavelet denoising and Kalman filtering noise reduction steps, time series differencing and sliding window feature dimension expansion steps, as well as EnKF boundary condition optimization and UDF dynamic pressure component compensation techniques, forming a comprehensive data quality improvement system. This process effectively counteracts external interference in test bench experiments, providing a reliable data foundation for model training and prediction, and is the core guarantee of anti-interference capability. Implementation support: Integrated architecture based on MATLAB / SIMULINK Scope of Protection: This protection integrates data acquisition, feature engineering, theoretical calculation, machine learning correction, and result output within the MATLAB / SIMULINK environment, enabling automated operation and real-time response of the prediction process. It also supports rapid re-prediction after dynamic adjustments to operating conditions during experiments. This architecture supports the engineering implementation of the core technical solution, ensuring the fulfillment of real-time requirements; it is an auxiliary but necessary protection measure.

[0093] The core objective of this invention is to provide a hybrid method for predicting pipeline flow characteristics that combines theoretical calculation with machine learning correction. This method ensures basic adaptability through mechanistic modeling and improves accuracy through data-driven correction. Specific objectives are as follows: The maximum relative error of pipeline flow parameter prediction is strictly controlled within 8% to ensure that the prediction results output by the model have high accuracy and reliability.

[0094] This invention needs to have strong anti-interference capabilities, effectively resist external noise interference, and ensure stable operation in complex and ever-changing working conditions, especially in the dynamic process of rapid flow changes in the transition state, to maintain the accuracy and adaptability of the prediction function.

[0095] The system should meet the requirements of real-time and high efficiency. The prediction response time under a single working condition must be completed within 5 seconds. It should also support rapid re-prediction based on dynamic adjustments during the test to effectively respond to the actual engineering needs of "real-time monitoring and immediate decision-making" in test bench testing, and improve the overall test efficiency and response capability.

Claims

1. A method for predicting the flow characteristics of a pipeline network in an environmental simulation system, characterized in that, Includes the following steps: S1. Combining theoretical derivation and CFD numerical simulation, we explore the evolution mechanism of pressure drop, outlet pressure and outlet temperature of the environmental simulation system under transient conditions. S2. Based on the investigation in S1, the flow velocity time series data of the pipeline outlet was collected through the transition state test data, and the calculation formulas and steps for the pipeline outlet pressure and temperature of different pipeline structures were derived based on the equations of mass conservation, momentum conservation and energy conservation. The collected flow velocity time series data was preprocessed. S3. Perform feature filtering on the preprocessed flow velocity time series data, and obtain key features by calculating the correlation between all filtered features and the outlet temperature correction coefficient and the outlet pressure correction coefficient. S4. Based on the derived calculation formulas and steps for pipeline outlet pressure and temperature, corrections are made in conjunction with transition state test data. A correction coefficient prediction model for outlet pressure and outlet temperature is established through machine learning algorithms. Key features are used as the input to the correction coefficient prediction model, and outlet temperature correction coefficient and outlet pressure correction coefficient are used as the outputs of the correction coefficient prediction model. S5. Use the correction coefficient prediction model to predict the flow characteristics of the pipeline network.

2. The method for predicting pipeline flow characteristics in an environmental simulation system according to claim 1, characterized in that, S3 is: Based on the preprocessed flow velocity time series data, key features related to the outlet temperature correction coefficient and the outlet pressure correction coefficient are screened. Combined with variable sensitivity analysis, a general feature system for cross-straight pipes, tees, two-stream flow systems, and exhaust diffusers is established to complete the unified feature screening across the pipeline network structure and obtain key features.

3. The method for predicting pipeline flow characteristics in an environmental simulation system according to claim 1, characterized in that, The prediction of pipeline flow characteristics using a correction coefficient prediction model includes: Using a multilayer perceptron (MLP) model, the first equivalent outlet pressure is obtained by predicting the outlet pressure of a typical pipeline network, in order to predict the flow characteristics of the pipeline network. By using the gradient boosting decision tree model GBDT for the exhaust diffuser cavity, the second equivalent outlet pressure and equivalent outlet temperature are obtained by predicting the outlet pressure / temperature of the exhaust diffuser, so as to predict the flow characteristics of the pipeline network.

4. The method for predicting pipeline flow characteristics in an environmental simulation system according to claim 3, characterized in that, In the multilayer perceptron (MLP) model, the calculation formulas and steps for pipeline outlet pressure under different pipeline network structures are derived as follows: Input pipe length Pipe inner diameter Local drag coefficient and pipeline inlet cross-sectional pressure ; Acquire time-series data of pipeline outlet flow velocity collected from transition state test data; The time-varying Reynolds number was calculated based on the flow velocity time series data. ; Based on time-varying Reynolds number Iterative solution of dynamic friction coefficient ; Based on dynamic friction coefficient Flow velocity time series data Pipe length Pipe inner diameter and local drag coefficient The total pressure difference of the pipeline under unsteady flow was calculated. ; Based on the pressure at the pipe inlet section Total pressure difference in unsteady flow of pipeline The pipeline outlet pressure was calculated. .

5. The method for predicting pipeline flow characteristics in an environmental simulation system according to claim 4, characterized in that, The formulas for calculating pipeline outlet pressure and temperature include: In a one-inlet-one-outlet pipe cavity model, the pipe outlet pressure The expressions for the temperature test values ​​are as follows: ； ； in, This indicates the pressure at the pipe inlet section. This represents the total pressure difference during unsteady flow in the pipeline. Indicates the dynamic friction coefficient. Indicates gas density, This indicates the temperature at the pipe outlet section. Indicates the pipe wall temperature. This indicates the temperature at the pipe inlet section. This represents the convective heat transfer coefficient of the gas. This represents the specific heat capacity of a gas at constant pressure; the coefficient of frictional resistance. The result is obtained through dynamic calculation using the following two formulas; When the Reynolds number is The calculation is performed using the Blasius formula: ； When the Reynolds number is The calculation is performed using the Nikolaiz formula: ； Time-varying Reynolds number , and instantaneous velocity related: ； In a one-inlet, two-outlet pipe cavity model, the pipe outlet pressure Pipeline outlet pressure The expressions for the temperature test values ​​are as follows: ； ； ； ； in, and These represent the pipeline outlets. and pipeline outlet The total pressure difference in unsteady flow, and These represent the pipeline outlets. and pipeline outlet The time-varying frictional resistance coefficient, and These represent the pipeline outlets. and pipeline outlet The corresponding pipe length, and These represent the pipeline outlets. and pipeline outlet The inner diameter of the pipe, and These represent the pipeline outlets. and pipeline outlet The local drag coefficient, and These represent the pipeline outlets. and pipeline outlet instantaneous speed, and These represent the pipeline outlets. and pipeline outlet The outlet temperature, and Indicates pipeline outlet and pipeline outlet The gas convection heat transfer coefficient; In a one-inlet, four-outlet pipe cavity model, the pipe outlet pressure The expressions for the temperature test values ​​are as follows: ； ； in, Indicates pipeline The total pressure difference in unsteady flow, Indicates pipeline The time-varying frictional resistance coefficient, Indicates pipeline Length, Indicates pipeline inner diameter, Indicates pipeline The local drag coefficient, Indicates pipeline instantaneous speed, Indicates pipeline The outlet temperature, Indicates the convective heat transfer coefficient at different pipe outlets , i Indicates pipeline inlet in and four exits out 1. out 2. out 3 and out The number 4 indicates the convective heat transfer coefficient. Obtained through dynamic calculation: ； in, Pr represents the thermal conductivity of the gas, and Pr represents the Prandtl number. In a two-inlet, one-outlet pipe cavity model, the pipe outlet pressure... The expressions for the temperature test values ​​are as follows: ； ； in, Indicates pipeline inlet The cross-sectional pressure, Indicates pipeline outlet The total pressure difference in unsteady flow, Indicates pipeline outlet The time-varying frictional resistance coefficient, Indicates pipeline outlet The length of the pipe, Indicates pipeline outlet The inner diameter of the pipe, Indicates pipeline outlet The time-varying frictional resistance coefficient, Indicates pipeline outlet instantaneous speed, This indicates the flow rate at the inlet of the first pipe. This indicates the inlet temperature of the first pipe. This indicates the flow rate at the inlet of the second pipe. This indicates the inlet temperature of the first pipe. k Indicates the first k One pipeline inlet, Indicates the outlet convective heat transfer coefficient. This indicates the average temperature of the inlet pipe.

6. The method for predicting pipeline flow characteristics in an environmental simulation system according to claim 3, characterized in that, The method of using a multilayer perceptron (MLP) model to predict the outlet pressure of a typical pipeline network and calculate the first equivalent outlet pressure includes the following steps: Extract key features from the flow velocity time series data under the new operating conditions; The extracted key features are input into a multilayer perceptron (MLP) model to predict the outlet pressure correction coefficient. By combining the derived formulas and steps for calculating pipeline outlet pressure and temperature, and substituting the outlet pressure correction coefficient, the first equivalent outlet pressure is calculated.

7. The method for predicting pipeline flow characteristics in an environmental simulation system according to claim 6, characterized in that, The expression for the first equivalent outlet pressure is as follows: ； ； in, This indicates the first equivalent export pressure. This represents the export pressure correction factor. Indicates the pipeline outlet pressure. This represents a multilayer perceptron. Indicates the valve opening and closing time. Indicates valve flow rate. Indicates valve temperature. Indicates pipe diameter. Indicates the valve opening and closing mode. Indicates the length of the pipe.

8. The method for predicting pipeline flow characteristics in an environmental simulation system according to claim 3, characterized in that, In the GBDT decision tree model for gradient boosting of the exhaust diffuser cavity, the calculation formula and steps for the pipeline outlet pressure of different pipeline network structures are derived as follows: Input import speed Imported static temperature T in Imported static pressure P in And to acquire time-series data of pipeline outlet flow velocity from transition state test data acquisition pipeline. ; Calculate the total inlet temperature T 0out and total pressure P 0out ; Based on the total temperature of the import T 0out and total pressure P 0out The total outlet temperature was calculated. T 0out and total pressure P 0out ; Analysis of corrected exhaust diffuser efficiency and total pressure recovery coefficient η p ; Based on the analysis results and the total outlet temperature T 0out and total pressure P 0out Calculate the outlet static temperature T out and imported static pressure P out : ； ； ； ； in, Indicates the export speed. Indicates the total import temperature. This represents the specific heat capacity of a gas at constant pressure. Indicates the import speed. This indicates the efficiency of the exhaust diffuser. k The adiabatic index indicates the thermal properties of a gas. η p This represents the total pressure recovery coefficient.

9. The method for predicting pipeline flow characteristics in an environmental simulation system according to claim 8, characterized in that, The exhaust diffuser cavity gradient boosting decision tree model GBDT, by predicting the exhaust diffuser outlet pressure / temperature, obtains the second equivalent outlet pressure and equivalent outlet temperature, including the following steps: Based on the derived calculation formulas and steps for pipeline outlet pressure of different pipeline network structures, corrections are made in conjunction with transition state test data; Based on the corrected results, the second equivalent outlet pressure is obtained using the gradient boosting decision tree model GBDT for the exhaust diffuser cavity. and equivalent outlet temperature : ； ； ； ； in, This represents the correction factor for the pipeline outlet pressure. Indicates imported static temperature. Indicates the import speed. Indicates the export speed. Indicates the time of mutation. Indicates the imported static pressure. This represents the correction factor for the outlet temperature of the pipeline network.

Images