An aircraft fleet structure load health monitoring method based on machine learning

Abstract

The application provides an aircraft fleet structure load health monitoring method based on machine learning, and belongs to the technical field of aircraft structure reliability and health monitoring. The application fuses single-aircraft full-parameter load spectrum measurement and machine learning modeling to perform aircraft fleet structure load prediction and life monitoring. The application is suitable for batch structure health state evaluation, damage diagnosis and residual life prediction of aircraft fleets of the same type and technical state. Through single-aircraft full-parameter measurement modeling and fleet small-parameter modification mode, the load prediction accuracy is improved while the cost is controlled, and accurate data support is provided for aircraft life extension.

Description

Technical Field

[0001] This invention belongs to the field of aircraft structural reliability and health monitoring technology, and particularly relates to a machine learning-based method for monitoring the structural load health of aircraft fleets. Background Technology

[0002] Flight load spectrum testing of aircraft is a core preliminary step in aircraft structural reliability research, serving as a crucial basis for conducting full-aircraft fatigue tests, determining structural life, and implementing life extension programs. Current aircraft load spectrum testing technology is relatively mature. The standard procedure involves selecting a single aircraft from existing or production-line aircraft of the same model, completing test modifications (including strain gauge and sensor placement), ground calibration tests, flight data acquisition, and subsequent data processing. Ultimately, the load spectrum test data of this single aircraft serves as the basis for fatigue life assessment of the entire aircraft model.

[0003] Meanwhile, aircraft structural health monitoring technology has become a research hotspot in the industry. Existing health monitoring technologies are mainly divided into two categories: one is based on finite flight state parameters (such as altitude, speed, and attitude angles) recorded by the aircraft flight control system, combined with aircraft design parameters, to estimate structural fatigue life through theoretical calculations or finite element simulations; the other introduces machine learning technology, integrating flight state parameters with usage data such as aircraft availability and maintenance records to achieve health monitoring of major structural loads. However, both methods have significant technical shortcomings, making it difficult to meet the needs of precise monitoring of large fleets.

[0004] In recent years, machine learning technology has been successfully applied in data mining and predictive modeling. Its technical path of "data acquisition - feature extraction - model training - predictive output" provides effective technical support for resolving the contradiction between accuracy and economy in aircraft fleet health monitoring. Based on this, there is an urgent need in this field for a health monitoring method that can achieve "full parameter measurement of a single aircraft + modification of a small number of parameters in the fleet + machine learning modeling," thereby improving the accuracy of aircraft fleet load monitoring while reducing testing costs by exploring the correlation between load and key parameters.

[0005] The existing technology has many shortcomings in practical applications and is difficult to meet the needs of batch health monitoring of aircraft fleets. The specific defects are as follows: (1) Insufficient representativeness of single aircraft test data: The existing technology uses the load spectrum test data of a single aircraft to represent the life of dozens or even hundreds of aircraft of the entire model, ignoring the load distribution differences caused by factors such as manufacturing tolerances and operating conditions of different aircraft, resulting in poor accuracy of life assessment results and failing to provide accurate health monitoring basis for each aircraft. (2) Extremely high cost of full aircraft load spectrum test: The modification of load spectrum test for a single aircraft requires the arrangement of hundreds to thousands of test channels, supporting ground calibration tests, flight tests and other links. The test cost for a single aircraft is as high as tens of millions of yuan and the cycle is as long as 6 to 10 months (1 to 2 years for large aircraft); if full aircraft test is carried out for a batch of aircraft, the cost and cycle will increase exponentially, which is not feasible in engineering. (3) Existing health monitoring methods have low accuracy: Theoretical calculations or finite element simulations based on the limited flight parameters of the flight control system are indirect load estimations with uncontrollable errors. Even with the introduction of machine learning technology, the load regression error is generally 5% to 10% because it only relies on flight parameters and non-direct load-related data (attendance rate, maintenance records, etc.) without combining the direct load signals of the structure. The error of non-primary load-bearing structures such as the vertical tail and landing gear is even more than 15%. (4) Insufficient data sampling rate leads to information loss: The sampling rate of the onboard flight parameter recording system is usually no more than 16Hz (mostly 4 to 8Hz), while the load spectrum measurement requires a sampling rate of 32 to 64Hz, and key structures such as the landing gear require 256Hz. The modeling method based solely on flight parameter data will result in the loss of a large amount of dynamic load information and will not be able to fully restore the load change characteristics during flight. Summary of the Invention

[0006] To address the issues of high testing costs for full-aircraft load spectrum measurements in existing technologies, insufficient representativeness of single-aircraft test data, and low regression accuracy, missing data, and large life prediction bias in existing health monitoring methods, this invention provides a structural load health monitoring method for aircraft fleets with controllable testing costs and high accuracy in structural load and life prediction. Through a technical approach of "full-parameter load spectrum measurement for a single aircraft + machine learning modeling + minor parameter modifications for the fleet," this method achieves accurate acquisition of loads on various structural components of each aircraft in a fleet of the same model, providing reliable data support for batch health monitoring and life extension of the fleet.

[0007] To achieve the above objectives, this invention discloses a machine learning-based method for monitoring the structural load health of an aircraft fleet, the method comprising: S1. Select one aircraft from the same type of in-service aircraft or production line aircraft for load spectrum flight test and modeling. S2. Conduct full-parameter testing and modification of the selected aircraft, including load spectrum testing and modification, strain testing and modification, and sensor testing and modification. S3. Conduct ground calibration tests on the modified aircraft to establish the relationship equation between the load and strain signal of each component; S4. Conduct load spectrum flight tests on the calibrated aircraft, and simultaneously collect onboard flight status parameters, modified sensor and strain parameters; the flight tests cover different aircraft attitudes, takeoff and landing weights, and flight airspace conditions. S5. Preprocess the collected flight status parameters and modified sensor parameters, substitute the relationship equation between the load and strain signal of each component into the preprocessed data, and calculate the actual load of each component; at the same time, substitute the sensor calibration coefficient to obtain the standard physical quantity data of each sensor. S6. Based on machine learning algorithms, perform correlation analysis on the preprocessed data and screen characteristic parameters that are strongly correlated with the load of each structural component. S7. Select the appropriate machine learning algorithm based on the load characteristics of different structural components; S8. Using flight parameters and characteristic parameters, a load regression prediction model is established using the stepwise reduction method; S9. Based on the installation parameters required by the load regression prediction model of each component, conduct channel testing and modification on other aircraft in the same model group. S10. Real-time collection of flight status parameters and added characteristic parameters for each modified aircraft in the fleet; S11. Substitute the collected data of each aircraft in the fleet into the load regression prediction model to calculate the actual load of each structural component of each aircraft, thereby realizing batch health monitoring of the structural load of the fleet.

[0008] In S1, the selected aircraft are the finalized aircraft, whose structure and expected usage are consistent with the fleet.

[0009] In S2, strain testing modification includes arranging strain gauges or fiber optic grating sensors at key structural locations, and sensor testing modification includes adding overload, displacement, temperature, and pressure sensors.

[0010] In S3, ground calibration tests are performed on the modified aircraft to collect loading signals and sensor strain signals of each structural component. The relationship equation between the load and strain signal of each component is established using the multiple linear regression method.

[0011] In S4, for aircraft models that do not support real-time acquisition of flight status parameters, the flight parameter data is copied from the ground and then time-aligned with the sensor data.

[0012] In S5, the preprocessing includes data validity judgment, multi-source signal time alignment, error identification and clearing, and zero drift correction.

[0013] In S6, the corresponding correlation analysis method is selected based on the data distribution characteristics. Specifically, this includes: Pearson correlation coefficient for quantitative data that conforms to a normal distribution, Spearman correlation coefficient for non-normally distributed data, and Kendall correlation coefficient for ordinal categorical variables. Combined with the importance ranking of model parameters, the root mean square error (RMSE) and coefficient of determination (R²) of the model are calculated. 2 As eigenvalues, the weight of each parameter's influence on the load prediction results is measured, resulting in an eigenvalue importance ranking table; the eigenvalues ​​cover flight state parameters, strain parameters, and other sensor parameters.

[0014] In S7: For components where the load and characteristic parameters are strongly linearly correlated, linear regression or multinomial regression models are selected; for the stress-bearing structures of wings and tail fins, random forest models are selected; for components subjected to high-frequency impact loads, neural network models are selected; the model type is adjusted according to the actual prediction results.

[0015] In S8, a progressive reduction method is used to establish a load regression prediction model, which includes: incorporating all characteristic parameters into the model and eliminating redundant parameters that are strongly correlated with each other; performing error analysis, parameter adjustment and progressive reduction based on the model regression prediction results, prioritizing the reduction of parameters of the added sensors, and finally determining the load regression prediction model composed of flight state parameters and added parameters.

[0016] In S9, the channel testing modification for other aircraft in the same model group includes: placing strain gauges or sensors with corresponding characteristic parameters in the same positions as the selected aircraft, and adding a small data acquisition system.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) Significantly improved load prediction accuracy: The existing load inversion method based on flight parameters has an error of 5%~10%, and the error of structures such as vertical tail and landing gear exceeds 15%; This invention uses direct correlation modeling of "flight parameters + a small number of added sensor parameters" to control the load regression error to 1%~3%, providing an accurate data basis for life prediction; (2) Improved efficiency of machine learning modeling: The introduction of sensor data that directly reflects the stress on the structure avoids the problems of low feature extraction efficiency and model getting trapped in local optima caused by relying solely on flight parameter data. The model can achieve rapid convergence in one go, and the modeling process does not require manual intervention, which greatly improves data processing efficiency. (3) The cost of aircraft group monitoring is greatly reduced: only one typical aircraft needs to be modified for full parameter measurement, and other aircraft in the group only need to be modified for a small number of channels. There is no need to repeat ground calibration tests; the test cycle and cost are only 5% to 10% of the full aircraft measurement, which solves the cost bottleneck of batch monitoring of aircraft groups. (4) Adapting to the batch monitoring needs of aircraft fleets: Through a low-cost modification scheme, the load of each aircraft in the same type of fleet can be accurately monitored, and the actual load usage of each aircraft can be obtained in batches, making up for the deficiency that the actual measurement of a single aircraft cannot cover the individual differences of the fleet. (5) Optimization of life prediction accuracy: According to the fatigue damage calculation theory of metal structure, a 10% load error will lead to a 50% life error; the load error of this invention is controlled within 3%, which can control the structural life prediction error within 10%, greatly improving the accuracy of fleet life assessment and providing a reliable basis for life determination and extension. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This invention illustrates a technical roadmap for monitoring the structural load health of aircraft fleets in an embodiment of the present invention. Figure 2 A flowchart illustrating the machine learning-based method for monitoring the structural load health of an aircraft fleet is shown in an embodiment of the present invention. Figure 3 This illustrates a logic diagram for constructing a machine learning model in an embodiment of the present invention. Figure 4 A schematic diagram comparing the modification parameters of typical aircraft and aircraft fleets in embodiments of the present invention is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Technical framework for monitoring the structural load health of aircraft fleets, such as Figure 1 As shown, the main process of a machine learning-based method for monitoring the structural load health of an aircraft fleet is as follows: Figure 2 As shown (see) Figure 2 (Steps S001-S010 in the process).

[0022] Typical aircraft selection: Select one representative aircraft (which must be a finalized aircraft with the same structure and expected usage as the fleet) from the existing or production line aircraft of the same type for load spectrum flight testing and modeling.

[0023] Load spectrum testing modification: Selected typical aircraft are modified for full-parameter testing, including but not limited to the placement of strain gauges (or fiber optic grating sensors) in key structural parts, and the addition of overload, displacement, temperature, and pressure sensors to ensure full coverage of the test channels and provide sufficient data support for subsequent feature parameter extraction and model building.

[0024] Ground calibration test: Ground calibration test is carried out on a typical modified aircraft to collect loading signals of each structural component and strain signals of the sensor. The relationship equation between the load and strain signal of each component is established by using the multiple linear regression method.

[0025] Load spectrum flight test: Load spectrum flight test is carried out on typical aircraft that have been calibrated, and flight status parameters, modified sensors and strain parameters are collected simultaneously. For aircraft that do not support real-time acquisition of flight status parameters, the flight parameter data is copied from the ground and then time-aligned with the sensor data. The flight test needs to cover typical operating conditions such as different aircraft attitudes and movements, takeoff and landing weights, and flight airspace to ensure the representativeness of the sample data.

[0026] Flight test data processing: The collected flight status parameters and modified sensor parameters are preprocessed, including data validity judgment, multi-source signal time alignment, error identification and clearing, zero drift correction, etc. The load-strain relationship equation established in step S3 is substituted into the preprocessed data to calculate the actual load of each component. At the same time, the calibration coefficients of temperature, pressure and other sensors are substituted to obtain the standard physical quantity data of various sensors.

[0027] Feature parameter extraction: Correlation analysis is performed on the preprocessed measured data using machine learning algorithms to screen feature parameters that are strongly correlated with the loads of each structural component; the corresponding correlation analysis method is selected according to the data distribution characteristics: Pearson correlation coefficient is used for quantitative data that conforms to a normal distribution, Spearman correlation coefficient is used for non-normally distributed data, and Kendall correlation coefficient is used for ordinal categorical variables; the importance of model parameters is ranked (by calculating the root mean square error (RMSE) and coefficient of determination (R²) of the model). 2 The eigenvalues ​​are used to measure the weight of each parameter's influence on the load prediction results, and finally, a ranking table of the importance of the eigenvalues ​​is obtained. The eigenvalues ​​cover flight state parameters, strain parameters, and other sensor parameters.

[0028] Machine learning model selection (e.g.) Figure 3As shown in the diagram, the labels are as follows: 21-Multi-source data input, 22-Correlation analysis unit, 23-Parameter importance ranking unit, 24-Model screening unit, 25-Stepwise reduction optimization unit, 26-High-precision model output): Based on the load characteristics of different structural components, select the appropriate machine learning algorithm: for components where the load and feature parameters are strongly linearly correlated, use linear regression or multinomial regression models; for complex stress structures such as wings and tail fins, use random forest models; for components such as landing gear subjected to high-frequency impact loads, use neural network models; the model type can also be adjusted according to the actual prediction results.

[0029] Establishment of a high-precision regression prediction model: The model is constructed using a stepwise reduction method. First, all highly correlated feature parameters selected in step S6 are included in the model, and redundant parameters with strong correlations are removed (such as the left / right flap angle, which is retained only). Error analysis, parameter adjustment, and stepwise reduction are performed based on the regression prediction results of the model. The parameters of the added sensors are removed first (the flight status parameters common to the fleet are retained). Finally, a high-precision load regression prediction model is determined, consisting of multiple flight status parameters and a small number of added parameters (the model of a single load parameter has no more than 3 added parameters).

[0030] Aircraft fleet modification: Based on the load prediction model of each component established in step S8, a small number of channel test modifications are carried out on other aircraft of the same model fleet. Strain gauges or sensors with corresponding characteristic parameters are placed in the same positions as typical aircraft, and a small data acquisition system is installed. The modification workload is only 10% to 20% of that of typical aircraft.

[0031] Flotation data acquisition: Real-time acquisition of flight status parameters and added characteristic parameters for each modified aircraft in the fleet.

[0032] Fleet load health monitoring: The collected data of each aircraft in the fleet are substituted into the high-precision regression prediction model established in step S8 to calculate the actual load of each structural component of each aircraft, thereby realizing batch health monitoring of the structural load of the fleet.

[0033] The first embodiment of the present invention (wing load health monitoring) will be described below.

[0034] Step S1: Select one currently serving aircraft of this model, whose flight hours and maintenance records are at the average level of the fleet to ensure representativeness.

[0035] Step S2: Arrange bending moment (M), shear force (Q), and torque (T) load test bridges on each test section of the wing. The number of bridges on each test section shall not be less than 6. At the same time, temperature sensors and overload sensors shall be installed on the leading edge and wingtip of the wing. The total number of test channels shall not be less than 200.

[0036] Step S3: Conduct ground calibration tests, set up individual calibration and composite calibration conditions for bending moment, shear force, and torque, repeat each condition 3 times, and use the multiple linear regression method to establish the relationship equation between each load parameter and the strain bridge signal to ensure that the load regression accuracy of the calibration equation is not less than 3%.

[0037] Step S4: Install the sensor with a sampling rate of 64Hz. During the flight test, synchronously collect onboard flight status parameters (altitude, speed, Mach number, dynamic pressure, static pressure, real-time fuel quantity, attitude angle, control surface angle, engine speed, center of gravity three-dimensional overload, etc.), and record the takeoff weight and payload weight. The flight conditions cover typical maneuvers such as level flight, climb, dive, and turn, and a total of no less than 20 flight tests are completed.

[0038] Step S5: Preprocess the collected data, align the flight parameter data and sensor data using timestamps, identify and remove abnormal errors using the 3σ criterion, and correct data zero drift using the moving average method; substitute the data into the ground calibration equation to calculate the actual loads M, Q, and T for each section.

[0039] Step S6: Use Pearson correlation coefficient (data verified to conform to normal distribution) for correlation analysis, and combine it with the parameter importance ranking of the random forest model to finally select 2 strain parameters (strain on the upper surface of the wing root and strain on the leading edge) and 8 flight parameter parameters (altitude, speed, Mach number, real-time fuel quantity, engine speed, center of gravity normal overload, flap angle, and aileron angle) as characteristic parameters for the load on each section of the wing.

[0040] Step S7: The wing load is significantly affected by airflow and structural coupling, so a random forest model is selected for modeling.

[0041] Step S8: The initial model incorporates all 9 feature parameters. After analysis, strongly correlated velocities and Mach numbers are removed (velocity parameters are retained). The final model input parameters are determined to be 2 strain parameters and 7 flight parameter parameters. The model MAPE is 1.9%.

[0042] Step S9: Modify other aircraft in the same fleet as the typical aircraft by placing two sets of strain gauges on the upper surface of the wing root and the leading edge, and adding a 9-channel data acquisition system. The modification cycle is only 15% of that of the typical aircraft.

[0043] Steps S10-S11: During flight of each modified aircraft, seven flight parameters and two strain parameters are collected simultaneously and substituted into the random forest prediction model to obtain the M, Q, and T loads of each wing section in real time, thereby achieving accurate monitoring of the wing load of the aircraft group.

[0044] The second embodiment of the present invention (landing gear load health monitoring) will be described below.

[0045] The implementation process was adjusted to take into account the characteristics of landing gear bearing high-frequency impact loads.

[0046] Step S2: Install high sampling rate strain sensors in key parts such as landing gear struts and rocker arms. At the same time, flight parameters such as brake pressure and landing gear retraction / extension signals need to be collected.

[0047] Step S4: Set the sensor sampling rate to 256Hz, and focus the flight test on conditions such as takeoff, landing, and ground taxiing.

[0048] Step S7: Select a neural network model (BP neural network) for modeling to meet the nonlinear prediction requirements of high-frequency impact loads.

[0049] Step S8: The final model input parameters are determined to be 2 added strain parameters (support strain, rocker arm strain) + 4 flight parameter parameters (normal overload, coasting speed, brake pressure, turning angle), and the model load regression error is controlled within 3%.

[0050] Step S9: The aircraft group retrofit only requires the installation of 2 sets of strain gauges, and the retrofit workload is 10% of that of a typical aircraft, achieving low-cost batch monitoring of landing gear loads.

[0051] In addition, this invention provides a comparison of modification parameters for typical aircraft and fleets, such as... Figure 4 As shown, the labels are explained as follows: 31 - Typical aircraft full parameter modification channel, 32 - Flotation limited parameter modification channel.

[0052] In summary, by analyzing the correlation of multi-source data and ranking the importance of parameters, the minimum set of feature parameters for load prediction is selected to achieve accurate modeling of "flight parameters + a small number of additional parameters"; a stepwise reduction method is used to build a machine learning model, prioritizing the retention of common flight parameters for the fleet and reducing additional parameters, thus realizing a technical solution for low-cost retrofitting and batch monitoring of the fleet; and a technical solution for adapting differentiated machine learning algorithms to the load characteristics of different structural components (random forest for wings / tails, neural networks for landing gear, etc.).

[0053] Compared with the prior art, the present invention has the following beneficial effects: (1) Significantly improved load prediction accuracy: The existing load inversion method based on flight parameters has an error of 5%~10%, and the error of structures such as vertical tail and landing gear exceeds 15%; This invention uses direct correlation modeling of "flight parameters + a small number of added sensor parameters" to control the load regression error to 1%~3%, providing an accurate data basis for life prediction; (2) Improved efficiency of machine learning modeling: The introduction of sensor data that directly reflects the stress on the structure avoids the problems of low feature extraction efficiency and model getting trapped in local optima caused by relying solely on flight parameter data. The model can achieve rapid convergence in one go, and the modeling process does not require manual intervention, which greatly improves data processing efficiency. (3) The cost of aircraft group monitoring is greatly reduced: only one typical aircraft needs to be modified for full parameter measurement, and other aircraft in the group only need to be modified for a small number of channels. There is no need to repeat ground calibration tests; the test cycle and cost are only 5% to 10% of the full aircraft measurement, which solves the cost bottleneck of batch monitoring of aircraft groups. (4) Adapting to the batch monitoring needs of aircraft fleets: Through a low-cost modification scheme, the load of each aircraft in the same type of fleet can be accurately monitored, and the actual load usage of each aircraft can be obtained in batches, making up for the deficiency that the actual measurement of a single aircraft cannot cover the individual differences of the fleet. (5) Optimization of life prediction accuracy: According to the fatigue damage calculation theory of metal structure, a 10% load error will lead to a 50% life error; the load error of this invention is controlled within 3%, which can control the structural life prediction error within 10%, greatly improving the accuracy of fleet life assessment and providing a reliable basis for life determination and extension.

[0054] Please note that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above embodiments only illustrate several implementations of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A method for monitoring the structural load health of an aircraft fleet based on machine learning, characterized in that, The method includes: S1. Select one aircraft from the same type of in-service aircraft or production line aircraft for load spectrum flight test and modeling. S2. Conduct full-parameter testing and modification of the selected aircraft, including load spectrum testing and modification, strain testing and modification, and sensor testing and modification. S3. Conduct ground calibration tests on the modified aircraft to establish the relationship equation between the load and strain signal of each component; S4. Conduct load spectrum flight tests on the calibrated aircraft, and simultaneously collect onboard flight status parameters, modified sensor and strain parameters; the flight tests cover different aircraft attitudes, takeoff and landing weights, and flight airspace conditions. S5. Preprocess the collected flight status parameters and modified sensor parameters, substitute the relationship equation between the load and strain signal of each component into the preprocessed data, and calculate the actual load of each component; at the same time, substitute the sensor calibration coefficient to obtain the standard physical quantity data of each sensor. S6. Based on machine learning algorithms, perform correlation analysis on the preprocessed data and screen characteristic parameters that are strongly correlated with the load of each structural component. S7. Select the appropriate machine learning algorithm based on the load characteristics of different structural components; S8. Using flight parameters and characteristic parameters, a load regression prediction model is established using the stepwise reduction method; S9. Based on the installation parameters required by the load regression prediction model of each component, conduct channel testing and modification on other aircraft in the same model group. S10. Real-time collection of flight status parameters and added characteristic parameters for each modified aircraft in the fleet; S11. Substitute the collected data of each aircraft in the fleet into the load regression prediction model to calculate the actual load of each structural component of each aircraft, thereby realizing batch health monitoring of the structural load of the fleet.

2. The method for monitoring the structural load health of an aircraft fleet based on machine learning according to claim 1, characterized in that, In S1, the selected aircraft are the finalized aircraft, whose structure and expected usage are consistent with the fleet.

3. The method for monitoring the structural load health of an aircraft fleet based on machine learning according to claim 2, characterized in that, In S2, strain testing modification includes arranging strain gauges or fiber optic grating sensors at key structural locations, and sensor testing modification includes adding overload, displacement, temperature, and pressure sensors.

4. The method for monitoring the structural load health of an aircraft fleet based on machine learning according to claim 3, characterized in that, In S3, ground calibration tests are performed on the modified aircraft to collect loading signals and sensor strain signals of each structural component. The relationship equation between the load and strain signal of each component is established using the multiple linear regression method.

5. The method for monitoring the structural load health of an aircraft fleet based on machine learning according to claim 4, characterized in that, In S4, for aircraft models that do not support real-time acquisition of flight status parameters, the flight parameter data is copied from the ground and then time-aligned with the sensor data.

6. The method for monitoring the structural load health of an aircraft fleet based on machine learning according to claim 5, characterized in that, In S5, the preprocessing includes data validity judgment, multi-source signal time alignment, error identification and clearing, and zero drift correction.

7. The method for monitoring the structural load health of an aircraft fleet based on machine learning according to claim 6, characterized in that, In S6, the corresponding correlation analysis method is selected based on the data distribution characteristics, specifically including: Pearson correlation coefficient is used for quantitative data that conforms to a normal distribution, Spearman correlation coefficient is used for non-normally distributed data, and Kendall correlation coefficient is used for ordinal categorical variables. Based on the importance ranking of model parameters, the root mean square error (RMSE) and coefficient of determination (R²) of the model are calculated. 2 As eigenvalues, the weight of each parameter's influence on the load prediction results is measured, resulting in an eigenvalue importance ranking table; the eigenvalues ​​cover flight state parameters, strain parameters, and other sensor parameters.

8. The method for monitoring the structural load health of an aircraft fleet based on machine learning according to claim 7, characterized in that, In S7: For components where the load and characteristic parameters are strongly linearly correlated, linear regression or multinomial regression models are selected; for the stress-bearing structures of wings and tail fins, random forest models are selected; for components subjected to high-frequency impact loads, neural network models are selected; the model type is adjusted according to the actual prediction results.

9. The method for monitoring the structural load health of an aircraft fleet based on machine learning according to claim 8, characterized in that, In S8, a progressive reduction method is used to establish a load regression prediction model, which includes: incorporating all characteristic parameters into the model and eliminating redundant parameters that are strongly correlated with each other; performing error analysis, parameter adjustment and progressive reduction based on the model regression prediction results, prioritizing the reduction of parameters of the added sensors, and finally determining the load regression prediction model composed of flight state parameters and added parameters.

10. The method for monitoring the structural load health of an aircraft fleet based on machine learning according to claim 9, characterized in that, In S9, the channel testing modification for other aircraft in the same model group includes: placing strain gauges or sensors with corresponding characteristic parameters in the same positions as the selected aircraft, and adding a small data acquisition system.

Images