A digital test method for aero-engine test run

Abstract

This invention relates to a digital testing method for aero-engine testing. The invention includes the following steps: S01) First, historical physical test data is collected; S02) A large digital test model M of the same type / structure is trained using the data from step S01); S03) Test data from the n-1 phases of physical testing of the engine under test is collected; S04) The data from the n-1 phases of physical testing of the engine under test is input into the large model from S02) and the model is fine-tuned to obtain the final large model M”; S05) The model accuracy is calculated; S06) The model accuracy is compared with a set threshold ε. If the model accuracy is less than the threshold, proceed to step S07), otherwise proceed to step S08); S07) The engine under test undergoes digital testing in the nth phase; S08) The engine under test undergoes physical testing in the nth phase. This invention reduces testing risks and resource investment while improving the efficiency of aero-engine manufacturing and maintenance, and shortening product delivery cycles.

Description

Technical Field

[0001] This invention mainly relates to the research fields of digital test of aero-engines, large models based on aero-engine test data, and transfer learning technology, and in particular to a digital test method for aero-engine testing. Background Technology

[0002] As the core power system of an aircraft, the performance and reliability of the engine directly determine the safety and economy of flight. Therefore, before manufacturing and maintenance, engines undergo multiple test runs to test and evaluate their performance and vibration conditions. During engine manufacturing and maintenance, a factory test run is conducted after the initial assembly of the production engine. Its main purpose is to check the manufacturing and assembly quality, as well as the operation of the engine and its accessories, and to adjust the engine and its accessories to achieve design performance. The inspection test run, conducted after the factory test run, a second disassembly and inspection, troubleshooting, and reassembly, aims to check the assembly quality under all operating conditions. Only when the engine performs satisfactorily and operates normally under all conditions can it be delivered to the customer.

[0003] Factory testing generally consists of three stages: break-in testing, adjustment testing, and preliminary operation testing. Break-in testing refers to the process of gradually increasing and then decreasing the mechanical, aerodynamic, and thermal loads on the assembled engine according to a prescribed testing procedure. Adjustment testing is conducted after break-in testing. This involves adjusting the engine's control components and modifying the control software to adjust key characteristics such as starting performance, thrust or power, fuel consumption rate, and acceleration / deceleration performance to meet performance requirements. It also involves adjusting the key characteristics of the engine's transmission, lubrication, and control systems to meet functional requirements. Preliminary operation testing is a preliminary check to ensure the engine meets delivery specifications. It is conducted according to the testing procedure approved by the user department and focuses on confirming key engine parameters including thrust or power, engine speed, fuel consumption rate, exhaust temperature, and mechanical vibration values. Generally, core measurement parameters during factory testing include: engine speed, fuel flow rate, thrust or power, temperature and pressure at characteristic sections of the engine's airflow channels, temperature and pressure in the lubrication and control systems, and mechanical vibration values.

[0004] For manufacturing and repair plants, inspection and testing generally include stages such as dry-running tests, idle-running tests, load tests, high-altitude tests, and extreme temperature tests. During the dry-running test, the engine is mounted on a test stand, and the rotor is driven by an electric motor for dry-running. This stage is to check whether the engine's mechanical components are operating normally, including various valves, pipes, oil lines, water lines, and other systems for leaks, oil leaks, and other problems. During the idle-running test, the engine is started but without external load. During the load test, the engine is mounted on a load stand, which simulates external loads for load testing. High-altitude tests refer to testing the performance and reliability of aero engines in high-altitude environments. Extreme temperature tests refer to testing aero engines in extreme temperature environments, such as high and low temperatures.

[0005] During engine manufacturing and maintenance, to ensure the safety and reliability of aircraft flights after delivery, physical test runs must be conducted on the assembled engines. However, due to the numerous components of an aero engine, assembly is often not completed in one go. Therefore, multiple test runs are conducted throughout the engine assembly and maintenance process, such as factory tests and inspection tests. Furthermore, to prepare for simulating various usage scenarios after engine delivery, each test run involves multiple test procedures / stages, thereby completing a comprehensive evaluation of the engine's post-assembly quality.

[0006] Traditional ground-based physical testing has several drawbacks due to the complex structure of aero engines and the harsh environment of testing operations:

[0007] 1) High risk: Ground physical testing needs to be conducted in a closed room or enclosed test site because engine testing may generate extremely high noise and temperature. Such an environment not only increases the difficulty of operation, but may also pose a health hazard to the staff.

[0008] 2) High cost: Ground physical testing requires a large amount of equipment and tools, including fuel and cooling systems, control consoles, sensors, and data acquisition systems. The purchase, maintenance, and upgrades of this equipment all require significant financial investment.

[0009] 3) Time-consuming and inefficient: Ground-based physical testing requires a significant amount of time and manpower, as it necessitates multiple tests and the analysis and comparison of test data. This can potentially impact the engine's delivery cycle and efficiency. Summary of the Invention

[0010] To address the aforementioned technical problems in the background art, this invention constructs a digital test model for aero-engines based on a machine learning method. Digital test runs are then conducted based on this model, thereby enabling digital assistance or replacement of some test procedures in ground physical test runs. This reduces test run risks and resource investment while improving the efficiency of aero-engine manufacturing and maintenance, and shortening product delivery cycles.

[0011] The technical solution of this invention is: This invention is a digital testing method for aircraft engine testing, characterized in that: the method includes the following steps:

[0012] S01) First, collect historical physical test data;

[0013] S02) Use the data from step S01) to train a large digital test model M of the same type / structure;

[0014] S03) Collect test data for the engine under test that has completed the n-1 stage physical test;

[0015] S04) Input the data from the n-1 stage physical test of the engine under test into the large model in S02) and fine-tune the model to obtain the final large model M”;

[0016] S05) Calculate the accuracy of the model;

[0017] S06) Compare the model accuracy with the set threshold ε. If the model accuracy is less than the threshold, proceed to step S07); otherwise, proceed to step S08.

[0018] S07) The engine under test will undergo digital test in the nth stage;

[0019] S08) The physical test of the engine under test is carried out in the nth stage.

[0020] Furthermore, the large digital test model M is mainly composed of an intake component network, a fan component network, a compressor component network, a combustion chamber component network, a turbine component network, and an exhaust nozzle component network. The characteristic parameters of each component are mainly composed of environmental parameters, state parameters, and control parameters during the test process.

[0021] Furthermore, the characteristic parameters corresponding to the intake duct network are intake duct parameters, such as intake pressure and temperature; the characteristic parameters corresponding to the fan component network are fan parameters, such as low-pressure fan speed; the characteristic parameters corresponding to the compressor component network are compressor parameters, such as compressor guide angle and high-pressure speed; the characteristic parameters corresponding to the combustion chamber component network are combustion chamber parameters, such as throttle lever position and angle; the characteristic parameters corresponding to the low-pressure turbine component network are low-pressure turbine component parameters, such as low-pressure turbine speed; the characteristic parameters corresponding to the low-pressure turbine component network are low-pressure turbine component parameters, such as high-pressure turbine speed; and the characteristic parameters corresponding to the exhaust nozzle component network are exhaust nozzle component parameters, such as exhaust nozzle diameter or exhaust nozzle distance.

[0022] Furthermore, the large digital test model M extracts the matching physical features of each component through the coupled network layer, the general operating law extraction layer is for learning and extracting the general operating features of the same model / structure engine in the whole test phase, and the mapping layer maps the test operation physical features to the target mapping parameters. The target mapping parameters are the target parameters of interest in the engine test test - performance, lubrication and vibration parameters.

[0023] Furthermore, fine-tuning yields the final large model "M". The first layer of this model is the target engine operation law extraction layer, which is used to extract the unique operation law features of the target engine under test. Its interface is the general operation law extraction layer of the large digital test model. The output of this layer flows to the mapping network layer. The input and output are the data from the physical test of the target engine. Its main function is to train and update the weights of the fine-tuning model network layers (target engine operation law feature extraction layer and fine-tuning model mapping layer). For the large model network layers, the weights of the network layers are frozen and directly called.

[0024] Furthermore, the following steps are included between steps S03) and S04):

[0025] S14) Substitute the physical test data of the engine under test in the first n-2 stages into the large model and fine-tune the large model to obtain the first updated large model M';

[0026] S15) Using the environmental and control data of stage n-1 as input parameters of M', calculate and predict the performance, lubricating oil and vibration parameters of stage n-1, and calculate the relative error ε' between the predicted value and the physical test value.

[0027] S16) Compare with the error threshold. If the relative error of all parameter predictions is within the threshold range or the mean of the relative error of all parameter predictions is within the threshold range, proceed to the next step S04); otherwise, proceed to S08.

[0028] Furthermore, in step S15), the environmental parameters are engine intake air temperature and intake air pressure; the control parameters are engine control parameters such as high and low pressure speeds, fan inlet guide vane angle, compressor inlet guide vane angle, throttle lever angle, and tailpipe diameter; the performance parameters mainly refer to engine performance parameters such as thrust, fuel consumption rate, and turbine exhaust temperature T6; the vibration parameters mainly refer to the horizontal and vertical vibration parameters of the engine intake casing, intermediate casing, accessory casing, and turbine rear casing. The formula for calculating the relative error of the prediction of performance, lubricating oil, and vibration parameters is:

[0029]

[0030] in denoted as predicted values ​​for performance, lubrication, and vibration parameters, and p represents the corresponding physical test values.

[0031] Furthermore, the historical physical test data in step S01) refers to the ground physical test data of m engines that have completed all stages of testing.

[0032] Furthermore, in step S02), "same model / same structure" refers to m engines having the same model or basically similar structure.

[0033] Furthermore, step S05) calculates the model accuracy using a common machine learning calculation method, such as the loss value in the final step of model training.

[0034] The beneficial effects of this invention are:

[0035] 1) First, this invention constructs a digital test model for aero-engines based on machine learning methods (consisting of a large digital test model containing general test operation characteristics and a fine-tuning model containing the specific operation rules of the target test engine). Based on this model, digital tests can be carried out during the production, manufacturing and maintenance of aero-engines, thereby enabling digital auxiliary tests or replacement tests for some test procedures in ground physical tests. This reduces test risks and resource input, while improving the efficiency of aero-engine manufacturing and maintenance and shortening the product delivery cycle.

[0036] 2) Furthermore, the model described in this invention uses physical test data from all completed test phases to train a large digital test model to obtain the general operating characteristics of a specific engine model or structure. It then fine-tunes the model using physical test data from the target engine under test and learns the engine's unique characteristics. This saves model training resources while enhancing the model's predictive generalization ability across all test phases / programs, avoiding the drawback of poor prediction accuracy due to not learning about the test phases to be conducted. (For example, assuming the engine under test has not yet undergone phase n of testing, but the large model has learned the general operating rules of phase n testing for m engines using the data from the engines that have already undergone testing, the engine under test can accurately predict its performance, lubrication, and vibration under the environmental, state, and control conditions of phase n testing by utilizing these general operating rules and the specific operating rules from the n-1 phases of testing it has already conducted.) Attached Figure Description

[0037] Figure 1 This is a flowchart of the digital test run process for an aero-engine according to the present invention;

[0038] Figure 2 This is a flowchart of the digital test run with prediction accuracy verification according to the present invention;

[0039] Figure 3 This is a structural diagram of the large-scale test model of the aero-engine and the fine-tuning model of the target engine of the present invention. Detailed Implementation

[0040] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments:

[0041] See Figure 1 To facilitate subsequent descriptions, the application scenario is described below: Assume that ground testing (primarily factory testing and inspection testing) during the manufacturing and maintenance of aero-engines involves n stages of testing / procedures (note that the number of testing stages may vary for different engine models, and the number of stages for factory testing and inspection testing also differs due to their different purposes). Furthermore, assume that all stages of ground physical testing for m engines have already been completed. For the engine under test, only n-1 stages of testing have been completed, while the nth stage has not been tested. The method described in this invention aims to conduct digital testing for the nth stage, thereby assisting existing physical testing or reducing the number of tests / time. Specifically, the process of the first specific embodiment of this invention is as follows:

[0042] Step S01) First, collect historical physical test data. Specifically, historical physical test data here refers to ground physical test data of m engines that have completed all phases of testing.

[0043] Step S02) Use the data from step S01 to train a large digital test model M of the same model / structure. Specifically, "same model / structure" here refers to m engines having the same model or basically similar structure.

[0044] Step S03) This step mainly involves collecting test data for the engine under test that has completed the n-1 stage of physical testing.

[0045] Step S04) Input the data from the completed physical test of the engine under test into the large model in S02) and fine-tune the model to obtain the final large model M”;

[0046] Steps S05) and S06) calculate the model accuracy and compare the model accuracy with the set threshold ε. If the model accuracy is less than the threshold, proceed to step S07); otherwise, proceed to step S08.

[0047] Step S07) The engine under test undergoes digital test in the nth stage;

[0048] Step S08) The physical test of the engine under test is carried out in the nth stage.

[0049] In step S02), the features learned by the large model refer to the common features of the engine under all test conditions (including n stages) for the same model or structure. The subsequent fine-tuning operation is to integrate the features of the engine under test into the large model, so that the large model has both the common test features of the engine under test and its own difference features, thus giving it better generalization ability.

[0050] In step S05), the calculation of model accuracy is a common calculation method in machine learning, such as the loss value in the final step of model training.

[0051] Example 2: Furthermore, the present invention addresses... Figure 1 The process also proposes a more rigorous test procedure, such as... Figure 2 As shown, the main difference is that a new layer of decision constraint is added to the credibility of the large model in the middle (i.e., the large model is fine-tuned twice). The specific steps are as follows:

[0052] Step S11) First, collect historical physical test data;

[0053] Step S12) Use the data from step S11) to train a large digital test model M of the same type / structure;

[0054] Step S13) Collect test data for the engine under test that has completed the n-1 stage of physical testing;

[0055] Step S14) Substitute the physical test data of the engine under test in the first n-2 stages into the large model and fine-tune the large model to obtain the first updated large model M';

[0056] Steps S15) and S16) use the environmental and control data of the (n-1)th stage as input parameters of M' to calculate and predict the performance, lubricating oil and vibration parameters of the (n-1)th stage, and calculate the relative error ε' between the predicted value and the physical test value, and compare it with the error threshold. If the relative error of all parameters is within the threshold range (or the mean of the relative error of all parameters is within the threshold range), then proceed to the next step; otherwise, proceed to S21).

[0057] Specifically, environmental parameters include engine intake air temperature and intake air pressure; control parameters include high and low pressure speeds, fan inlet guide vane angle, compressor inlet guide vane angle, throttle lever angle, and tailpipe diameter; performance parameters mainly refer to engine performance parameters such as thrust, fuel consumption rate, and turbine exhaust temperature (T6); and vibration parameters mainly refer to the horizontal and vertical vibration parameters of the engine intake casing, intermediate casing, accessory casing, and turbine rear casing (the same applies below).

[0058] Furthermore, the formulas for calculating the relative errors in the predictions of performance, lubricating oil, and vibration parameters are as follows:

[0059]

[0060] in denoted as predicted values ​​for performance, lubrication, and vibration parameters, and p represents the corresponding physical test values.

[0061] Step S17: Substitute the physical test data of the engine under test in stage n-1 into the large model and fine-tune the large model to obtain the large model M” after the second fine-tuning update;

[0062] In particular, fine-tuning methods belong to the transfer learning techniques in machine learning. For example, by freezing all layers outside the fully connected layers of the model, only updating the weights of the original fully connected layers or updating the weights of the new fully connected layers, the transfer of learned features can be achieved.

[0063] Steps S18) and S19) calculate the model accuracy and compare the model accuracy with the set threshold ε. If the model accuracy is less than the threshold, proceed to step S20); otherwise, proceed to step S21.

[0064] Step S20) The engine under test undergoes a digital test run in the nth stage;

[0065] Step S21) The physical test of the engine under test is carried out in the nth stage.

[0066] Example 3: Construction and Training Method of Large-Scale Digital Test Model for Aero-engines

[0067] like Figure 3 The diagram shown is a structural diagram of the large-scale digital test model and its fine-tuning model of the aero-engine described in this invention. First, the large-scale digital test model is mainly composed of an inlet component network, a fan component network, a compressor component network, a combustion chamber component network, a turbine component network, and a tail nozzle component network. The characteristic parameters of each component are mainly composed of environmental parameters, state parameters, and control parameters during the test.

[0068] The characteristic parameters corresponding to the intake duct network are intake duct parameters, such as intake pressure and temperature; the characteristic parameters corresponding to the fan component network are fan parameters, such as low-pressure fan speed; the characteristic parameters corresponding to the compressor component network are compressor parameters, such as compressor guide angle and high-pressure speed; the characteristic parameters corresponding to the combustion chamber component network are combustion chamber parameters, such as throttle lever position and angle; the characteristic parameters corresponding to the low-pressure turbine component network are low-pressure turbine component parameters, such as low-pressure turbine speed; the characteristic parameters corresponding to the high-pressure turbine component network are low-pressure turbine component parameters, such as high-pressure turbine speed; the characteristic parameters corresponding to the exhaust nozzle component network are exhaust nozzle component parameters, such as exhaust nozzle diameter or exhaust nozzle distance. The coupling network layer extracts the matching operational physical features of each component, the general operational law extraction layer learns and extracts the general operational features of engines of the same model / structure throughout the entire test run, and the mapping layer maps the test run physical features to target mapping parameters. Specifically, the target mapping parameters of this invention are the target parameters of interest in engine test runs—performance, lubrication, and vibration parameters. Furthermore, for the fine-tuning model, its first layer is a target engine operation law extraction layer. This layer is used to extract the unique operation law features of the target engine under test. Its interface is the general operation law extraction layer of the large digital test model, and the output of this layer flows to the mapping network layer. In particular, the network proposed in this invention preferably uses neural networks such as RNN, LSTM, and GRU, as well as multi-head attention networks.

[0069] Finally, regarding the model training process, for the training of the large digital test model M, the model structure adopts... Figure 3 The left half of the model uses ground physical test data (environmental parameters, state parameters, and control parameters) from all stages of the m engines as input (i.e., characteristic parameter data of components such as the air intake, at the bottom of the model) and outputs target mapping parameters (performance, lubricating oil, and vibration parameters).

[0070] For the fine-tuning model, its input and output are the data from the physical test of the target engine. It mainly trains and updates the weights of the fine-tuning model network layers (the target engine operation law feature extraction layer and the fine-tuning model mapping layer). For the large model network layer on the left, its network layer weights are frozen and directly called.

[0071] The technical contents of this invention and those not specifically described in the above embodiments are the same as those in the prior art.

[0072] The present invention is not limited to the above embodiments; all embodiments described herein can be implemented and have the aforementioned good effects.

Claims

1. A digital testing method for aircraft engine testing, characterized in that: The method includes the following steps: S01) First, collect historical physical test data; S02) Use the data from step S01) to train a large digital test model M of the same type / structure; S03) Collect test data for the engine under test that has completed the n-1 stage physical test; S14) Substitute the physical test data of the engine under test in the first n-2 stages into the large model and fine-tune the large model to obtain the first updated large model M'; S15) Using the environmental and control data of stage n-1 as input parameters of M', calculate and predict the performance, lubricating oil, and vibration parameters of stage n-1, and calculate the relative error ε' between the predicted values ​​and the physical test values; the environmental parameters are engine intake air temperature and intake air pressure; the control parameters are high and low pressure speeds, fan inlet guide vane angle, compressor inlet guide vane angle, throttle lever angle, and tailpipe diameter; the performance parameters refer to thrust, fuel consumption rate, and turbine exhaust temperature T6; the vibration parameters include the horizontal and vertical vibration parameters of the engine intake casing, intermediate casing, accessory casing, and turbine rear casing. The formula for calculating the relative error of the predicted performance, lubricating oil, and vibration parameters is: ； in denoted as predicted values ​​for performance, lubrication, and vibration parameters, and p represents the corresponding physical test values. S16) Compare with the error threshold. If the relative error of all parameter predictions is within the threshold range or the mean of the relative error of all parameter predictions is within the threshold range, proceed to the next step S04); otherwise proceed to S08. S04) Substitute the data from the n-1 phase physical test of the engine under test into the large model in S02) and fine-tune the model to obtain the final large model M''; The first layer of the fine-tuned final large model M'' is the target engine operation law extraction layer, which is used to extract the unique operation law features of the target engine under test. Its interface is the general operation law extraction layer of the digital test large model. The output of this layer flows to the mapping network layer; the input and output are the data from the physical test of the target engine, which are used to train and update the weights of the fine-tuning model network layer. For the large model network layer, its network layer weights are frozen and directly called; S05) Calculate the accuracy of the model; S06) Compare the model accuracy with the set threshold ε. If the model accuracy is less than the threshold, proceed to step S07; otherwise, proceed to step S08. S07) The engine under test will undergo digital test in the nth stage; S08) The physical test of the engine under test is carried out in the nth stage.

2. The digital testing method for aircraft engine testing according to claim 1, characterized in that: The large digital test model M consists of an intake component network, a fan component network, a compressor component network, a combustion chamber component network, a turbine component network, and an exhaust nozzle component network. The characteristic parameters of each component are composed of environmental parameters, state parameters, and control parameters during the test.

3. The digital testing method for aircraft engine testing according to claim 2, characterized in that: The characteristic parameters corresponding to the intake duct component network are intake duct parameters, including intake pressure and temperature; the characteristic parameters corresponding to the fan component network are fan parameters, including low-pressure fan speed; the characteristic parameters corresponding to the compressor component network are compressor parameters, including compressor guide angle and high-pressure speed; the characteristic parameters corresponding to the combustion chamber component network are combustion chamber parameters, including throttle lever position and angle; the characteristic parameters corresponding to the low-pressure turbine component network are low-pressure turbine component parameters, including low-pressure turbine speed; and the characteristic parameters corresponding to the exhaust nozzle component network are exhaust nozzle component parameters, including exhaust nozzle diameter or exhaust nozzle distance.

4. The digital testing method for aircraft engine testing according to claim 3, characterized in that: The digital test model M extracts the matching physical features of each component through a coupled network layer. The general operation law extraction layer learns and extracts the general operation features of the same type / structure engine throughout the entire test phase. The mapping layer maps the test operation physical features to target mapping parameters, which are the target parameters of interest in the engine test—performance, lubrication, and vibration parameters.

5. The digital testing method for aircraft engine testing according to any one of claims 1 to 4, characterized in that: The historical physical test data in step S01 refers to the ground physical test data of m engines that have completed all stages of testing.

6. The digital testing method for aircraft engine testing according to claim 5, characterized in that: In step S02), "same model / same structure" means that m engines have the same model or the same structure.

7. The digital testing method for aircraft engine testing according to claim 6, characterized in that: The calculation of model accuracy in step S05) is a common calculation method in machine learning, specifically the loss value of the final step in model training.

Images