A modeling method for aircraft engine interpolation model based on machine learning

Abstract

The present invention discloses a method for modeling an aero-engine interpolation model based on machine learning, comprising the following steps: collecting aero-engine flight data and corresponding key parameters, performing feature selection, and preprocessing the raw data to establish a flight data set; using the converted input features as node coordinates and the converted output features as target values, searching the flight data set, preliminarily establishing an interpolation table, and marking missing nodes in the interpolation table; using samples from the flight data set to establish a machine learning model with the above converted features as input and output; and using the machine learning model to predict missing nodes in the interpolation table to obtain a complete interpolation table. The present invention avoids the problem of reduced accuracy caused by the traditional interpolation model being limited by the operating range of the flight data. Instead, the present invention utilizes the powerful generalization performance of random forests to establish a complete and accurate interpolation table. In actual use, the accuracy of the aero-engine interpolation model can be effectively improved.

Description

Technical Field

[0001] The present invention relates to modeling of key parameters of an aero-engine, and in particular to a modeling method of an aero-engine interpolation model based on machine learning. Background Art

[0002] With the continuous improvement of modern aircraft and engine performance, various control methods have emerged. Afterburner fuel control, anti-surge control, and other control variables have introduced increasing demands on engine control systems. Since the advent of the full authority digital electronic controller in 1970, due to its simple structure, light weight, good adaptability, and high reliability, it has quickly replaced traditional hydraulic systems and has been widely used in civil and military aircraft engines. The development of such complex and advanced aircraft engine CNC systems requires the use of aircraft engine mathematical models instead of real engines, from initial control law analysis and design to later semi-physical simulation tests. This reduces the number of tests, avoids test risks, and reduces costs. The stability and reliability of aircraft engine mathematical models are crucial to the development and design of aircraft engine control systems.

[0003] The main methods for modeling key parameters of aircraft engines include analytical and experimental methods. The analytical method is based on the aerodynamic thermodynamic laws followed during aircraft engine operation, utilizing relevant theorems and laws, and is established mathematically. It can calculate the state parameters of each engine section with good accuracy. However, in the absence of detailed engine component characteristics, establishing an accurate engine analytical model is very difficult. The experimental method does not consider the complex internal mechanisms of the system. It is based on processing engine test data to obtain its characteristics, thereby obtaining a mathematical model. Domestic scholars often establish simplified models based on engine flight data to simulate the engine's steady-state and dynamic processes in real time. Zhou Wenxiang used the dynamic coefficient method to establish a simplified real-time model of a dual-shaft turbofan engine, while Feng Haifeng used engine component-level models to obtain the engine's operating characteristic lines and established the engine model using the characteristic line method. The interpolation model is a model established using the experimental method. Because it does not rely on the characteristics of engine components and has good accuracy and real-time performance, it is widely used in semi-physical simulation tests of engine numerical control systems. Its main disadvantage is that due to the limitations of flight conditions and the influence of mechanical structure, thermal factors and other factors, it is difficult to obtain accurate flight data under all operating conditions, resulting in a decrease in the accuracy of the model under some operating conditions, which hinders the implementation of control simulation tests.

[0004] Currently, many researchers at home and abroad are using machine learning algorithms to study aircraft engine performance parameter prediction models and have achieved relevant research results. S. Sina et al. used dynamic neural networks to detect aircraft engine performance status faults, achieving a prediction error rate of less than 5%. L. J. Kerr et al. proposed a real-time estimation model for aircraft engine damage based on the Kalman filter algorithm. Xu Jianxin et al. proposed an ensemble of extreme learning machines based on the AdaBoost.RT ensemble algorithm, demonstrating that the improved model's prediction performance outperformed other models. Wang Xuhui et al. proposed a least squares support vector machine algorithm based on adaptive grid search optimization to establish low-altitude data link communication with enroute aircraft and predict gas path parameter trends. Cui Jianguo combined a grey prediction model with a relevance vector machine prediction model, overcoming the shortcomings of grey prediction, reducing data randomness, and improving the accuracy of the prediction model. Pi Jun et al. used a generalized regression neural network to accurately predict exhaust temperature. Zhang Yizhen et al. proposed a combined method based on dynamic weighted kernel density estimation, using an improved ensemble learning algorithm to improve the accuracy of aircraft engine prediction models. Currently, there is a significant research gap in the application of machine learning algorithms to aircraft engine interpolation modeling. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a modeling method for an aircraft engine interpolation model based on machine learning, thereby improving the accuracy of the interpolation model within the full flight envelope and enhancing the real-time monitoring of the working status of key engine parameters.

[0006] Technical solution: The present invention provides a method for modeling an aircraft engine interpolation model based on machine learning, comprising the following steps:

[0007] (1) Collect the flight data of the aircraft engine and the corresponding key parameters, perform feature selection, and preprocess the raw data to establish a flight data set, which mainly includes the following contents:

[0008] (1.1) Based on the aerodynamic thermodynamics laws followed during the operation of an aircraft engine, the inlet pressure, temperature, and given fuel quantity are extracted as input features; the temperature and pressure of each section in the compressor and turbine are extracted as output features.

[0009] (1.2) Using the similarity principle, preprocess the extracted features, calculate the feature conversion coefficient, obtain the converted input features and output features, and establish the flight data set.

[0010] (2) The converted input features are used as node coordinates, and the converted output features are used as target values. The flight data set is searched, an interpolation table is preliminarily established, and the missing nodes in the interpolation table are marked.

[0011] (3) Using the samples in the flight data set, a machine learning model is established with the above converted features as input and output.

[0012] (4) Taking the missing nodes obtained in step (2) as input, the machine learning model is used to predict the missing nodes of the interpolation table, obtain a complete interpolation table, and establish an aircraft engine interpolation model.

[0013] The missing nodes mentioned in step (4) include the following:

[0014] Missing value coordinates:

[0015]

[0016] Where X is the reduced input feature, M is the dimension of the interpolation table, and N is the number of missing values;

[0017] The coordinate Z is used as the input of the random forest model, and the random forest model is used to calculate the output features of the node.

[0018] (5) Collect the input features of the flight data, preprocess them, use the converted input features as input, perform linear interpolation in the interpolation table, correct the interpolation results, and obtain the predicted values of the output features. Use the mean square error criterion to evaluate the accuracy of the model.

[0019] Beneficial effects: Compared with the existing technology, the present invention has the following advantages: the present invention uses the similarity principle to explore the intrinsic connection between features, obtains the interpolation table based on the reduced features, and then uses the machine learning algorithm to train the model. The generalization ability of the machine learning model is used to predict the missing values of the interpolation table, thereby improving the accuracy of the interpolation model within the full flight envelope and enhancing the real-time monitoring of the working status of key engine parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the steps of the present invention;

[0021] Figure 2 The distribution diagram of engine sensor data; 2a is the original data of the engine sensor, and 2b is the pre-processed flight data set;

[0022] Figure 3 Construct process diagrams for regression decision trees;

[0023] Figure 4 This is the random forest model structure diagram;

[0024] Figure 5It is the test engine prediction effect curve; among them, 5a is the parameter prediction effect of test data 1, 5b is the parameter prediction effect of test data 2, 5c is the parameter prediction effect of test data 3, and 5d is the parameter prediction effect of test data 4. DETAILED DESCRIPTION

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0026] like Figure 1 As shown, a modeling method of an aircraft engine interpolation model based on machine learning includes the following steps:

[0027] (1) Collect the flight data of the aircraft engine and the corresponding key parameters, perform feature selection, and preprocess the raw data to establish a flight data set, which mainly includes the following contents:

[0028] (1.1) Based on the aerodynamic thermodynamics laws followed during the operation of an aircraft engine, the inlet pressure, temperature, and given fuel quantity are extracted as input features; the temperature and pressure of each section in the compressor and turbine are extracted as output features.

[0029] (1.2) Using the similarity principle, preprocess the extracted features, calculate the feature conversion coefficient, obtain the converted input features and output features, and establish the flight data set.

[0030] (2) The converted input features are used as node coordinates, and the converted output features are used as target values. The flight data set is searched, an interpolation table is preliminarily established, and the missing nodes in the interpolation table are marked.

[0031] (3) Using the samples in the flight data set, a machine learning model is established with the above converted features as input and output.

[0032] (4) Taking the missing nodes obtained in step (2) as input, the machine learning model is used to predict the missing nodes of the interpolation table, obtain a complete interpolation table, and establish an aircraft engine interpolation model.

[0033] (5) Collect the input features of the flight data, preprocess them, use the converted input features as input, perform linear interpolation in the interpolation table, correct the interpolation results, and obtain the predicted values of the output features. Use the mean square error criterion to evaluate the accuracy of the model.

[0034] This case uses the flight data of an aircraft engine. The selected input features are: atmospheric pressure P0, inlet temperature T1, and a given fuel quantity W. fm , the output characteristic is the low-pressure turbine inlet temperature T 45 A two-dimensional interpolation table is proposed. The machine learning algorithm used is random forest.

[0035] Step 1: According to the aerodynamic thermodynamics laws followed during the operation of the aircraft engine, extract: atmospheric pressure P0, intake inlet temperature T1, given fuel quantity W fm As input feature; extract the low-pressure turbine inlet temperature T 45 As the output feature, the raw data is preprocessed. The preprocessing formula is defined as follows:

[0036] Temperature conversion factor:

[0037] Pressure conversion factor:

[0038] Equivalent fuel volume:

[0039] Reduced temperature at low-pressure turbine inlet:

[0040] After preprocessing the data, the intake duct inlet temperature T1 is selected and the fuel quantity W is converted to fmc is the reduced input characteristic, the low-pressure turbine inlet reduced temperature T 45c The output features are folded and randomly numbered to form the flight data set. The normalized original data and flight data set are as follows Figure 2 shown.

[0041] Step 2: Compare the intake duct inlet temperature T1 with the equivalent fuel quantity W fmc As the node coordinate, the low-pressure turbine inlet reduced temperature T 45c As the target value, a uniformly distributed two-dimensional grid is established. The size of the two-dimensional grid is m×n, and the structure is shown in Table 1. Due to the range of the flight dataset, some nodes in the grid are missing.

[0042] Table 1 Two-dimensional grid structure

[0043]

[0044] Step 3: The construction of decision tree is the basis of random forest model. First, the input is the intake inlet temperature T1 and the equivalent fuel volume W. fmc , the output is the low-pressure turbine inlet reduced temperature T 45c A single decision tree.

[0045] In the input space where the flight data set is located, X and Y are input samples and output samples respectively, forming a training data set D = {(x1, y1), (x2, y2), ..., (x N ,y N )}.in is the characteristic vector of the input instance, which is composed of the intake port inlet temperature T1 and the equivalent fuel quantity W fmcComposition, y i is an output example, namely the low-pressure turbine inlet reduced temperature T 45c , i=1,2,…,N, where N is the sample capacity.

[0046] Assume that the input space R has been divided into M subspaces, and there is a fixed output value in each subspace Then the regression decision tree model can be expressed as

[0047]

[0048] Where I is the pulse function; is the average value of the output variable y of all samples in each subspace, that is,

[0049]

[0050] In the subspace R corresponding to the current parent node m In the example, select the jth feature variable x (j) and its value s, define its two regions R1(j,s)={x|x (j) ≤s} and R2(j,s)={x|x (j) >s}, N1 and N2 are the number of samples in sub-regions R1 and R2 respectively. m For all the values of all the features in , the best one is selected as the split point (j, s) according to the square error minimization criterion. That is, solve the following formula

[0051]

[0052] in

[0053] According to the above division method, R1 and R2 are respectively used as parent nodes, and the division is performed recursively until the square error of the y value of the current sample set is less than the predetermined threshold. The division is stopped and the leaf node is forced to be generated. At this point, a single regression tree is established.

[0054] Decision tree, the construction process of decision tree is as follows Figure 3 shown.

[0055] Based on a single decision tree, we apply the concept of ensemble learning to construct multiple decision trees for training. Using bootstrap sampling, we extract M sample sets with replacement from the flight data set and construct M regression decision trees for training. The unselected samples are treated as out-of-bag data and form the test sample data for the model.

[0056] The flight dataset has p features. At each internal node of each decision tree model, k (k ≤ p) variables are randomly selected as candidate branch features. The optimal branch is found according to the process of building a single decision tree. Each decision tree branches recursively from top to bottom until the variance of the y-values of the samples in the current parent node is less than a given variance threshold. When this condition is met, the recursion stops and the current parent node is set as a leaf node.

[0057] According to the attribute characteristics of the data, the generated M decision trees generate a random forest prediction model according to the following rules.

[0058] y RF =ave(y(X,T a )),a=1,2,…,M

[0059] Where T a is the a-th regression tree; y(X,T a ) is the predicted value of the a-th regression tree. RF The predicted value is the average of the predicted values of M trees. The model structure of random forest is as follows Figure 4 shown.

[0060] Step 4: Set of node coordinates with missing values Where T1 is the inlet temperature of the node, W fmc is the equivalent fuel quantity of the node, and N is the number of nodes. i The predicted low-pressure turbine front-reduced temperature at that node is used as input to the random forest model. Predictions are performed for all missing nodes and the predicted values are populated into the grid nodes to obtain a uniformly distributed two-dimensional interpolation table, thus establishing an aircraft engine interpolation model.

[0061] Step 5: To evaluate the accuracy of the aircraft engine interpolation model based on random forest, extract the inlet inlet temperature T1 and the equivalent fuel volume W from the flight data set. fmc As input, the low-pressure turbine front reduced temperature T is obtained from the two-dimensional interpolation table using the linear interpolation method. 45c Calculate the temperature conversion coefficient K t , the reduced temperature before the low-pressure turbine T 45c Correction is performed to obtain the low-pressure turbine inlet temperature T 45 The final predicted value is as follows:

[0062]

[0063] T 45 =T 45c ·K t ·K t -273.15

[0064] The root mean squared error (RMSE) criterion is used, which is defined as follows:

[0065]

[0066] Where T represents the number of samples in the test set, Indicates the temperature before the low-pressure turbine T 45 The error between the predicted value and the actual value.

[0067] The effect obtained by the present invention is compared with the interpolation model before correction, and the comparison results are shown in Table 2. From the comparison of the results, it can be seen that the method provided by the present invention achieves excellent results.

[0068] In order to intuitively show the performance of the aircraft engine interpolation model based on random forest, this experiment selected flight data of four test engines and estimated them based on the interpolation model of random forest. Figure 5 As shown, the low-pressure turbine inlet temperature T 45 The predicted value is very consistent with the actual value and has high accuracy. Therefore, the method provided by the present invention can improve the accuracy of the interpolation model.

[0069] Table 2 Comparison of model prediction effects

[0070]

Claims

1. A method for modeling an aircraft engine interpolation model based on machine learning, characterized in that: The following steps are involved: (1) Collect flight data of aircraft engines and corresponding key parameters, perform feature selection, and preprocess the raw data to establish a flight data set; specifically: (1.1) Based on the aerodynamic thermodynamics laws followed during aircraft engine operation, the inlet pressure, temperature, and given fuel quantity are extracted as input features; the temperature and pressure of each section in the compressor and turbine are extracted as output features; (1.2) Using the similarity principle, preprocess the extracted features, calculate the feature conversion coefficient, obtain the converted input features and the converted output features, and establish the flight data set; (2) Using the converted input features as node coordinates and the converted output features as target values, search the flight data set, preliminarily establish an interpolation table, and mark the missing nodes in the interpolation table; (3) Using the samples in the flight data set, a machine learning model with the above-mentioned converted features as input and output is established, wherein the machine learning model is a random forest model; (4) Use the machine learning model to predict the missing nodes of the interpolation table, obtain the complete interpolation table, and establish the aircraft engine interpolation model.

2. The method for modeling an aircraft engine interpolation model based on machine learning according to claim 1, characterized in that: The missing nodes mentioned in step (4) include the following: Missing value coordinates: Where X is the reduced input feature, M is the dimension of the interpolation table, and N is the number of missing values; The coordinate Z is used as the input of the random forest model, and the random forest model is used to calculate the output features of the node.

3. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the aircraft engine interpolation modeling method based on machine learning according to any one of claims 1-2.

4. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for modeling an aircraft engine interpolation model based on machine learning according to any one of claims 1 to 2 is implemented.

Images