A data enhancement method of an aero-engine digital twin model
By setting the working state interval in the digital twin model of the aero-engine, calculating the proportion of target samples, and using generative adversarial networks to generate simulated data, the overfitting and underfitting problems caused by data inhomogeneity are solved, and the generalization ability and accuracy of the model are improved.
Patent Information
- Application Number
- CN202310991205.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-08
- Publication Date
- 2025-11-28
- Estimated Expiration
- 2043-08-08
AI Technical Summary
Existing digital twin models for aero-engines cannot account for the non-uniformity of actual operating data, leading to overfitting or underfitting under certain conditions, thus failing to meet the accuracy requirements for performance prediction.
By setting the judgment intervals for each operating state of the engine, calculating the proportion of target samples for each state, adjusting the data sample size, and using generative adversarial networks to generate simulated data, the data is ensured to be evenly distributed and data augmentation is performed.
It improves the model's generalization ability and robustness under various working conditions, enhances performance prediction accuracy, and prevents overfitting and underfitting.
Smart Images
Figure CN117113001B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of data preprocessing of aero-engine digital twin model, in particular to a data enhancement method of aero-engine digital twin model. BACKGROUND
[0002] The running state of an aero-engine directly affects flight safety, so it is necessary to monitor the real-time running state of the engine. In order to realize the real-time monitoring of some performance parameters that cannot be directly measured during flight, the mapping relationship between the unmeasurable performance parameters and the on-board measurable operating parameters can be established based on ground test data, and the unmeasurable performance parameters can be calculated in real time to realize the real-time monitoring of these performance parameters. The accuracy of the mapping relationship between the performance parameters and the measurable operating parameters directly affects the accuracy and reliability of performance monitoring and state judgment.
[0003] At present, the aero-engine digital twin modeling method based on deep learning technology has achieved high accuracy. The deep learning model can learn the mapping relationship between the performance parameters of the engine and the on-board measurable operating parameters through a large amount of data, so as to predict the performance of the engine under different working conditions.
[0004] By selecting appropriate algorithms, a theoretically high-precision aero-engine digital twin model can be established, and the prediction ability of the model for engine performance parameters is mainly derived from training data. The model learns the actual running data of the engine to find the mapping relationship between the measured performance parameters and other operating parameters. However, the data distribution in the actual running data of the engine is extremely uneven, and the running data under some working conditions accounts for a very small proportion, such as the acceleration and deceleration process accounts for a very small proportion, the proportion of the afterburner state is very small, and the proportion of some state data is even less than 1%. Most of the data are running data of the engine under stable maximum or throttling state. Due to the average effect, the model will repeatedly train the parameter response relationship under the state with a large proportion of data during model training, which is prone to overfitting. The model learns the response relationship between parameters under the state with a small proportion of data, which is prone to underfitting, and the model cannot accurately map the parameter response relationship under these states, resulting in a large local error in predicting the performance of the engine, even the maximum error exceeds the allowable range, and the accuracy cannot meet the engineering requirements, so the model cannot be used.
[0005] It is difficult to consider the unevenness of the actual data distribution by only selecting the algorithm and designing the model architecture, in order to improve the model precision, it is necessary to reconstruct and enhance the original operation data of the aero-engine before model training, improve the proportion of training data in some states, make the model learn the data response relationship of the aero-engine in each working state more comprehensively, and improve the generalization ability and overall robustness of the model.
[0006] The current aero-engine digital twin data preprocessing method mainly includes the following steps: ① selecting parameters in the original aero-engine operation data; ② cleaning the original data, that is, removing outliers; ③ normalizing the cleaned data, and then inputting it into the constructed digital twin model to start iterative training.
[0007] Since the data preprocessing process does not consider the state distribution of the actual operation data, the data distribution of each state is extremely uneven, and when training the aero-engine digital twin model, since the average error of each batch of data is used as the deviation for adjusting the model parameters, the average effect is easy to eliminate the part of the feature with small data proportion, the model learns insufficient features, the generalization ability and robustness are insufficient, and the precision of the digital twin model in predicting the performance of the engine is low. SUMMARY
[0008] The main purpose of the present application is to provide a data enhancement method for an aero-engine digital twin model, so as to solve the technical problem that the performance prediction precision cannot meet the requirements due to the inability to consider the unevenness of the original data distribution in the prior art.
[0009] The technical solution of the present application is: the present application is a data enhancement method for an aero-engine digital twin model, which is characterized in that: the method comprises the following steps:
[0010] 1) Obtain the actual operation data of the aero-engine, select the required parameters, and perform data cleaning;
[0011] 2) Based on the throttle lever position data obtained in step 1), set the judgment interval of each working state of the engine;
[0012] 3) According to the engine state judgment interval in step 2), count the proportion of each state in the original engine data;
[0013] 4) According to the engine state judgment interval in step 2), calculate the target sample proportion of each state data;
[0014] 5) Calculate the amount of data samples that need to be adjusted for each state;
[0015] 6) Divide the engine working states that need to be deleted and the engine working states that need to be added;
[0016] 7) deleting data samples;
[0017] 8) adding data samples;
[0018] 9) randomly shuffling all samples, completing the data enhancement of the aero-engine digital twin model training data samples.
[0019] Further, the specific steps of step 1) are: according to the aero-engine digital twin model established according to actual needs, selecting each parameter data and throttle lever position data required by the training model, and eliminating abnormal values in the data.
[0020] Further, the specific steps of step 2) are: according to the engine design standard, obtaining the overall change interval [a, b] of the throttle lever in the whole process, and setting the throttle lever change interval of each working state of the engine: start state [a1, b1], slow speed state [a2, b2], throttle state [a3, b3], maximum state [a4, b4], minimum afterburning state [a5, b5], and full afterburning state [a6, b6].
[0021] Further, the specific steps of step 3) are:
[0022] Obtain the total amount of engine original data samples: m, and obtain the original sample amount of the six engine working states in step two: n1, n2, n3, n4, n5, n6, using the following formula:
[0023] p i =(n i / m)·100%
[0024] Calculate the original sample proportion of each state: p1, p2, p3, p4, p5, p6.
[0025] Further, the specific steps of step 4) are: based on the overall change interval length of the throttle lever position in step 2) and the judgment interval length of each working state, using the following formula:
[0026] pt i =((b i -a i ) / (b-a))·100%
[0027] Calculate the target sample proportion of each state: pt1, pt2, pt3, pt4, pt5, pt6.
[0028] Further, the specific steps of step 5) are: based on the original sample proportion of each state calculated in step 3), the total amount of original data, the original sample amount of each state, and the target sample proportion of each state data calculated in step 4), using the following formula:
[0029] nt i =pt i ·m
[0030] Calculate the target data sample size of each state: nt1, nt2, nt3, nt4, nt5, nt6;
[0031] Then calculate the data sample size that needs to be adjusted for each state: dnt1, dnt2, dnt3, dnt4, dnt5, dnt6;
[0032] dnt i =int(nt i -n i )
[0033] Where: int() operator represents the integer operation.
[0034] Further, the specific steps of step 6) are: according to the data sample size that needs to be adjusted for each state calculated in step 5), if dnt i is less than 0, it means that part of the data in the state needs to be deleted, reducing the sample size; if dnt i is greater than 0, it means that the data in the state needs to be increased, increasing the sample size.
[0035] Further, the specific steps of step 7) are: for the working state that needs to be reduced in step 6), random deletion is performed, that is, in the original data sample of the state, first number all the samples in order, then generate a specified number dnt i of random numbers, and delete the samples with these numbers.
[0036] Further, the specific steps of step 8) are: for the working state that needs to be increased in step 6), sample increase is performed, and there are two optional sample increase methods:
[0037] Method one: randomly copy a specified number dnt i of original sample data of the state, that is, first number all the samples in order, then generate a specified number dnt i of random numbers, and add these numbered samples to the state sample again;
[0038] Method two: generate simulated samples of the state based on generative adversarial networks (GAN) and add them to the state sample, that is, first construct the corresponding generative adversarial network model based on the original sample data of each state that needs to be increased in step six, train to obtain the corresponding simulated data generator of each state, and then use the simulated data generator of the corresponding state to randomly generate a specified number dnti The data samples of the corresponding state are added into the samples of the corresponding state.
[0039] Further, the specific step of step 9) is that after step 7) and step 8) are completed, all samples of each state are combined, and all samples are randomly shuffled, so that the samples of each state are randomly and uniformly distributed, that is, the data enhancement of the training data samples of the aero-engine digital twin model is completed.
[0040] The aero-engine digital twin model data enhancement method has the following beneficial effects:
[0041] The aero-engine digital twin model data enhancement method can improve the extremely uneven distribution of state data samples in the original operation data of the aero-engine, so that the training data of the aero-engine digital twin model is randomly and uniformly distributed under each working state of the engine, accelerates the convergence of the aero-engine digital twin model training process, prevents overfitting in some states and underfitting in other states, improves the model generalization ability and robustness, and improves the model precision. BRIEF DESCRIPTION OF DRAWINGS
[0042] Figure 1 is a flowchart of an aero-engine digital twin model data enhancement method according to an embodiment of the present application;
[0043] Figure 2 is a flowchart of an aero-engine digital twin model data enhancement method according to an embodiment of the present application;
[0044] Figure 3 is a flowchart of an aero-engine digital twin model data enhancement method according to an embodiment of the present application;
[0045] Figure 4 is a flowchart of an aero-engine digital twin model data enhancement method according to an embodiment of the present application; DETAILED DESCRIPTION
[0046] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments:
[0047] Referring to Figure 1 , the aero-engine digital twin model data enhancement method according to the present application has the following process:
[0048] 1) Obtain the actual operation data of the aero-engine, filter the required parameters, and perform data cleaning;
[0049] 2) Based on the throttle lever position data obtained in step 1), set the judgment interval of each working state of the engine;
[0050] 3) According to the engine state judgment interval in step 2), the proportion of each state in the original engine data is counted;
[0051] 4) According to the engine state judgment interval in step 2), the target sample proportion of each state data is calculated;
[0052] 5) The data sample amount of each state that needs to be adjusted is calculated;
[0053] 6) The engine working states that need to delete data and need to add data are divided;
[0054] 7) The data sample is reduced;
[0055] 8) The data sample is increased;
[0056] 9) Randomly shuffle all samples to complete the data enhancement of the training data sample of the aero-engine digital twin model.
[0057] The following describes the implementation process of the data enhancement of the aero-engine thrust digital twin model as a specific embodiment of the present application:
[0058] Referring to Figure 2 , the training data of the aero-engine thrust digital twin model is taken as an example to perform data enhancement on the training data, and the specific implementation details are as follows:
[0059] Step 1) Obtain the ground test data of a certain type of aero-engine, and assume that it contains 100,000 data samples. The environmental parameters (including: environmental pressure, environmental temperature, engine inlet total temperature), control parameters (including: fan inlet guide vane angle, high-pressure compressor inlet guide vane angle, low-pressure rotor speed, high-pressure rotor speed, throttle lever position, oil pressure, tail nozzle throat diameter), and performance parameters (including: low-pressure turbine exhaust gas temperature, thrust) are selected out, and the data outliers are cleaned and removed.
[0060] Step 2) Based on the throttle lever position data obtained in step 1), set the judgment interval of each working state of the engine. According to the design standard of the engine, the overall change interval of the throttle lever is [0, 100], and the throttle lever change interval of each working state of the engine is set: start state [0, 10], slow state [10, 20], throttle state [20, 60], maximum state [60, 80], minimum afterburning state [80, 90], and full afterburning state [90, 100].
[0061] Step 3) According to the engine state judgment interval in step 2), the proportion of each state in the engine original data is counted. The total number of original data samples is m = 100000, and the original sample numbers of the six engine working states in step 2) are n1 = 2000, n2 = 3000, n3 = 20000, n4 = 70000, n5 = 4000, and n6 = 1000. The original sample proportions of each state are p1 = 2%, p2 = 3%, p3 = 20%, p4 = 70%, p5 = 4%, and p6 = 1%.
[0062] Step 4) The target sample proportion of each state data is calculated. Based on the total change interval length of the accelerator position in step 2) and the judgment interval length of each working state, the target sample proportion of each state is calculated using the following formula: pt1 = 10%, pt2 = 10%, pt3 = 40%, pt4 = 20%, pt5 = 10%, and pt6 = 10%.
[0063] Step 5) The data sample amount that needs to be adjusted for each state is calculated. Based on the original sample proportion of each state calculated in step 3), the total sample amount of the original data, the original sample amount of each state, and the target sample proportion of each state data calculated in step 4), the target data sample amount of each state is calculated using the following formula: nt1 = 10000, nt2 = 10000, nt3 = 40000, nt4 = 20000, nt5 = 10000, and nt6 = 10000. Then the data sample amount that needs to be adjusted for each state is calculated: dnt1 = 8000, dnt2 = 7000, dnt3 = 20000, dnt4 = -50000, dnt5 = 6000, and dnt6 = 9000.
[0064] Step 6) Divide the engine working states that need to delete data and add data. According to the data sample amount that needs to be adjusted for each state calculated in step 5), dnt4 is less than 0, indicating that part of the data in the maximum state of the engine needs to be deleted to reduce the sample amount; and the dnt of the remaining working states is greater than 0, indicating that data needs to be added in the remaining states to increase the sample amount of these states.
[0065] Step 7) Delete data samples. The original data samples in the maximum state of the engine in step 6) are randomly deleted, i.e. in the original data samples of this state, all samples are sequentially numbered from 0 to 70000, and then 50000 random numbers are generated in 0 to 70000, and the samples with these sequence numbers in the original data of the maximum state are deleted.
[0066] Step 8), increase the data sample. For the working state that needs to increase the sample in step 6), the sample increase is carried out, and in the embodiment, the simulation data samples of the engine in the starting state, the slow vehicle state, the throttling state, the minimum boost state and the full boost state are generated based on the generative adversarial network and added to the original data samples of the corresponding state.
[0067] As shown in Figure 3 , a flowchart for generating simulation data samples of each state using a generative adversarial network, that is, first, based on the original sample data of each state that needs to increase the sample, the corresponding generative adversarial network model is constructed. An optional model structure is shown in Figure 4 , the simulation data generator of each state (including: engine starting state simulation data generator, engine slow vehicle state simulation data generator, engine throttling state simulation data generator, engine minimum boost state simulation data generator, engine full boost state simulation data generator) is trained, and then a specified number dnt i of data samples are randomly generated using the simulation data generator of the corresponding state and added to the samples of the corresponding state.
[0068] Step 9), randomly shuffle all samples. After completing steps 7) and 8), all samples of each state are merged and all samples are randomly shuffled to randomly and uniformly distribute the samples of each state, that is, the data augmentation of the training data samples of the aero-engine digital twin model is completed.
[0069] The specific parameters involved in the present application are as follows:
[0070] T0: ambient temperature, P0: ambient pressure, T1: engine inlet total temperature, a1: fan inlet guide vane angle, a2: high-pressure compressor inlet guide vane angle, n1: low-pressure rotor speed, n2: high-pressure rotor speed, PLA: throttle lever position, Pm: oil pressure, D8: tail nozzle throat diameter, T6: low-pressure turbine afterburner gas temperature, F: thrust.
[0071] The technical content not specifically described in the present application and the above embodiments is the same as the prior art.
[0072] The present application is not limited to the above embodiments, and all the contents described in the present application can be implemented and have the good effects described.
[0073] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. A data augmentation method for an aeroengine digital twin model, characterized in that: The method comprises the following steps: 1) Obtain the actual operation data of the aero-engine, screen the required parameters, and perform data cleaning; The specific steps are: according to the aero-engine digital twin model established according to actual needs, select the parameter data and throttle lever position data required by the training model, and eliminate abnormal values in the data; 2) Based on the throttle lever position data obtained in step 1), set the judgment interval of each working state of the engine; The specific steps are: according to the engine design standard, obtain the overall change interval [a, b] of the throttle lever in the whole process, and set the throttle lever change interval of each working state of the engine: start state [a1, b1], slow state [a2, b2], throttle state [a3, b3], maximum state [a4, b4], minimum afterburning state [a5, b5], and full afterburning state [a6, b6]; 3) According to the engine state judgment interval in step 2), count the proportion of each state in the original engine data; The specific steps are: obtain the total amount of engine original data samples: m, and obtain the original sample amount of the six engine working states in step two: n1, n2, n3, n4, n5, n6, and use the following formula: p i = (n i / m) · 100% Calculate the original sample proportion of each state: p1, p2, p3, p4, p5, p6; 4) According to the engine state judgment interval in step 2), calculate the target sample proportion of each state data; The specific steps are: based on the total change interval length of the throttle lever position and the judgment interval length of each working state in step 2), use the following formula: pt i = ((b i -a i ) / (b-a)) · 100% Calculate the target sample proportion of each state: pt1, pt2, pt3, pt4, pt5, pt6; 5) Calculate the data sample amount that needs to be adjusted for each state; The specific steps are: based on the original sample proportion of each state calculated in step 3), the total sample amount of the original data, the original sample amount of each state, and the target sample proportion of each state data calculated in step 4), use the following formula: nt i = pt i · m Calculate the target data sample amount of each state: nt1, nt2, nt3, nt4, nt5, nt6; Then calculate the data sample amount that needs to be adjusted for each state: dnt1, dnt2, dnt3, dnt4, dnt5, dnt6; dnt i = int(nt i -n i ) Where: int() operator represents integer operation; 6) Divide the engine working states that need to delete data and need to add data; The specific steps are: according to the data sample quantity required to be adjusted in each state calculated in step 5), if dnt i is less than 0, it is indicated that part of data in the state needs to be deleted to reduce the sample quantity; if dnt i is greater than 0, it is indicated that data needs to be added in the state to increase the sample quantity; 7) Delete data samples; 8) Increase data samples; 9) Randomly shuffle all samples to complete the data enhancement of the aero-engine digital twin model training data samples.
2. The method of claim 1, wherein: The specific step of step 7) is: in the working state of step 6) requiring to delete samples, randomly deleting, that is, in the original data samples of the state, first numbering all samples in sequence, then generating random numbers of specified quantity dnt i , and deleting the samples of the numbers.
3. The method of claim 2, wherein: The specific steps of step 8) are: for the working states that need to add samples in step 6), add samples, and there are two optional sample addition methods: Method one: randomly duplicate the specified number dnt of the state original sample data, that is, first number the original sample in order, then generate a random number dnt of the specified number, and add these number samples to the state sample again; i Method two: randomly duplicate the specified number dnt of the state original sample data, that is, first number the original sample in order, then generate a random number dnt of the specified number, and add these number samples to the state sample again; i Method three: randomly duplicate the specified number dnt of the state original sample data, that is, first number the original Method two: generate the state simulation sample based on the generative adversarial network (GAN), and add the state sample, that is, first, based on the original sample data of each state of the sample to be increased in step six, construct the corresponding generative adversarial network model, train to obtain the simulation data generator corresponding to each state, and then use the simulation data generator corresponding to each state to randomly generate a specified number dnt i of data samples, and add them to the sample corresponding to the state.
4. The method of claim 3, wherein: The specific steps of step 9) are: after completing steps 7) and 8), combine all samples of each state, and randomly shuffle all samples to make the samples of each state randomly and uniformly distributed, that is, the data enhancement of the aero-engine digital twin model training data samples is completed.