Aircraft flight subject identification method based on time sequence flight parameter data

Abstract

The invention belongs to the field of aircraft structure fatigue life design, and particularly relates to an aircraft flight subject identification method based on time sequence flight parameter data. According to the method, time, external weight, whole aircraft weight, gravity center normal overload and gravity center lateral overload processes are extracted from multivariate time sequence flight parameter data, and error correction, filtering and compression are completed; based on correlation analysis between a physical base and flight parameters, flight parameter data strongly associated with gravity center overload are mined; performing feature extraction on the flight parameter data, and constructing a flight subject identification key feature database; and finally, machine learning models such as a support vector machine and a random forest are adopted to construct a flight subject prediction model, and an aircraft flight subject identification technology based on the time sequence flight parameter data is realized. According to the identification method provided by the invention, the blank of intelligently judging the flight subject is filled, and the labor cost is saved; the method is high in recognition efficiency and accurate, and can effectively avoid the problems of wrong marking and missing marking of manual marking of flight subjects.

Description

Technical Field

[0001] The invention belongs to the field of aircraft structure fatigue life design, and in particular relates to an aircraft flight subject identification method based on time-series flight parameter data. Background Art

[0002] During the design phase of an aircraft, a design load spectrum reflecting the expected use is compiled based on the design mission profile, and durability and damage tolerance analysis and tests are carried out based on this. Since the design load spectrum is difficult to truly reflect the use of the aircraft during the service phase, it is necessary to compile a measured load spectrum and revise the analysis and test results during the design phase to ensure the flight safety of the aircraft in service.

[0003] GJB67.6A "Repetitive Loads, Durability and Damage Tolerance" requires that when full-scale damage tolerance tests are carried out after strength test flights or batch production, the test load spectrum should, as far as possible, consider the correction of the design measured load spectrum by the measured load data.

[0004] The measured load spectrum must represent the overall usage of the fleet of in-service aircraft. When compiling the measured load spectrum, it is necessary to consider the proportion of take-offs and landings of each flight subject in the total number of take-offs and landings when the fleet is in use, and then combine it with the typical take-off and landing load data of each flight subject. Therefore, obtaining the proportion of take-offs and landings of each flight subject in the total number of flight take-offs and landings is crucial for compiling the measured load spectrum.

[0005] The current method of obtaining the proportion of flight subjects in a fleet mainly relies on ground crew to manually mark the flight subject data after takeoff and landing. This work needs to be carried out after each flight of each aircraft. In addition to increasing a large amount of manual workload, manual marking of flight subjects is also prone to problems such as mislabeling and missing labels. To solve this problem, it is necessary to invent a subject recognition technology based on flight parameters to realize automatic identification of flight subjects, so as to compile a more accurate measured load spectrum and provide support for more accurate life determination of field aircraft. Summary of the invention

[0006] Purpose of the invention: The present invention proposes a flight subject identification method, which extracts the characteristic values ​​of time-series flight parameter data, constructs a model for predicting flight subjects from flight parameter characteristics, and implements a flight subject identification method based on flight parameter characteristic data. The present invention fills the gap in intelligent identification of flight subjects, supplements and improves flight missions without recorded subjects, and ensures the integrity of the data used. It is efficient and accurate, avoiding problems such as mislabeling and missing labels that are prone to occur when manually marking flight subjects. This patent can support the compilation of aircraft measured load spectra, provide support for field aircraft fatigue life determination and ensure the safety of aircraft use.

[0007] Technical solution of the present invention

[0008] A method for identifying aircraft flight subjects based on time-series flight parameter data, the method comprising the following steps:

[0009] Step 1: Flight parameter data screening and cleaning: Initially, the default retention time, external weight, weight, center of gravity normal overload and center of gravity lateral overload are retained; by conducting correlation analysis between the physical basis and the flight parameters, multicollinearity analysis and stepwise regression are used for the remaining flight parameters to screen out the flight parameters with a high correlation with the center of gravity normal overload and the center of gravity lateral overload. The final 14 flight parameters screened include: time, external weight, weight, center of gravity normal overload, center of gravity lateral overload, altitude, Mach number, lift rate, roll angular velocity, pitch angular velocity, yaw angular velocity, bank angle, pitch angle, and heading angle. Then, the error codes, out-of-bounds data, and ground segment data of the flight parameters are eliminated to reduce the overall scale of the data;

[0010] Step 2: Data processing: Except for the time item, the flight parameter data processed in step 1 are used to generate new parameter items according to the physical meaning, mathematical and / or mechanical models of the flight parameters, such as flight altitude × Mach number, Mach number × rudder position signal, rudder position signal × left outer elevator position signal and other nonlinear items; then, multicollinearity diagnosis and filtering methods are used to reduce the scale of the sample space (i.e., dimensionality reduction processing) and standardize the data;

[0011] Step 3: Extracting features from time series data: Extract features from the time series flight parameter data processed in step 2 and build a flight parameter feature database; then combine the information of each take-off and landing flight subject to build a typical flight parameter feature & flight subject database;

[0012] Step 4: Using typical flight parameter features and flight subject database as input, build a flight subject prediction model based on time series flight parameter data to realize flight subject identification.

[0013] Furthermore, in step 3, when extracting features, the features of "time" are taken as take-off time, landing time and total flight time, the features of "weight" are taken as take-off weight and landing weight, and the features of "external weight" are taken as take-off external weight and landing external weight; for the remaining flight parameters except time, weight and external weight, the features are taken as their maximum value, minimum value, standard deviation and mean value. And the normal overload of the center of gravity and the weight of the whole aircraft are used as input to calculate the equivalent damage and nominal damage of the take-off and landing, which are also used as the feature data of the take-off and landing.

[0014] Furthermore, in step one, the bit errors and out-of-bounds data of each take-off and landing flight parameter are eliminated, small range filtering and peak filtering are performed; and the flight parameter data is identified as the air segment data by the flight altitude, the ground segment data is eliminated, and only the flight parameter data of the air segment is retained, thereby reducing the input data scale.

[0015] Furthermore, through the time-series flight parameter data, different methods are adopted to realize the extraction of flight parameter history data characteristics for different flight parameters, and based on the fatigue damage algorithm, the equivalent damage and nominal damage of the take-off and landing are calculated to realize the overall technical process of flight subject prediction.

[0016] Furthermore, when selecting flight parameters for the flight subject identification model in step 1, prior knowledge is used to preliminarily screen the flight parameter data to determine the initial selected flight parameter data. The regression analysis method is combined to finally determine the input flight parameters of the flight subject identification model.

[0017] Furthermore, in order to enable the machine learning model to effectively learn the product features between the input independent variables, the flying parameter data product term described in step 2 is used to enhance the data. And considering the possibility of strong linear correlation between the data after data enhancement, which will cause unnecessary increase in model training time, multicollinearity diagnosis is carried out to retain one of the multiple groups of data with strong linear correlation, so as to achieve dimensionality reduction of the model input independent variables, improve the stability of the model, and reduce the overall input data scale. The peak-to-valley filtering algorithm and the small-range filtering algorithm can be further used to filter the model input data based on the center of gravity overload, further reducing the scale of model training input data.

[0018] Furthermore, the data enhancement method described in step 2 multiplies the input flight parameter data by two or more to obtain product terms between the flight parameters, so that the machine learning model can learn the product features between the input value variables;

[0019] Furthermore, the multicollinearity diagnosis described in step 2 removes data with strong linear correlation between input independent variables, thereby avoiding model instability caused by multicollinearity of input data, achieving data dimensionality reduction, and avoiding unnecessary waste of resources in model training.

[0020] Furthermore, the filtering method described in step 2 takes the center of gravity normal overload as the filtering object, and adopts small variable range filtering to filter out the small overload cycles and intermediate data points in the cycles in various flight maneuvers; and adopts the peak-to-valley value filtering algorithm to filter out the flight parameter data cycles with small peak values ​​and large valley values, further reducing the scale of input data and speeding up the model training speed.

[0021] Furthermore, the "nominal damage" and "equivalent damage" features described in step three are due to the fact that there are certain differences in damage between different flight subjects, and the damage differences between takeoffs and landings of the same subject are small, while the damage differences between takeoffs and landings of different subjects are large. Therefore, a fatigue damage algorithm is used to calculate the nominal damage based on the normal overload history of the center of gravity, and the equivalent overload history is calculated using the normal overload history of the center of gravity of the takeoff and landing and the weight history of the entire aircraft, and the equivalent damage is calculated. "Equivalent damage" and "nominal damage" are also used as key features of the takeoff and landing.

[0022] Furthermore, the method described in step 4 is that the essence of subject identification is the classification between different categories of data. Combined with prior knowledge, typical machine learning models suitable for classification algorithms are adopted, such as support vector machine classification algorithm, random forest and other models, to build a model that uses time series flight parameter feature values ​​as input to predict flight subjects.

[0023] Beneficial effects of the present invention

[0024] The present invention establishes a flight subject identification method based on the characteristics of time-series flight parameter data. The method fills the gap in intelligent identification of flight subjects and saves labor costs. Compared with the method of marking flight subjects, the problem of missing flight subjects can be avoided. The method realizes automatic identification of flight subjects, provides strong support for the compilation of aircraft measured load spectra, field aircraft fatigue life determination and aircraft safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The present invention is a flow chart of a method for identifying aircraft flight subjects based on time-series flight parameter data;

[0026] Figure 2 Input flight parameter screening flow chart for flight subject identification model;

[0027] Figure 3 It is the feature extraction of time series flight parameters. Specific embodiments

[0028] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] Inspired by the above technical ideas, the present invention proposes a method for identifying aircraft flight subjects based on time-series flight parameter data in a specific implementation process, including the following implementation steps:

[0030] Step 1: Screening of multivariate time series flight parameter data.

[0031] S1.1: Initial screening of flight parameter data, using prior knowledge to make preliminary judgments, retaining the time, external load, total aircraft weight, center of gravity normal overload and center of gravity lateral overload data that are closely related to the flight subject;

[0032] S1.2: Carry out correlation analysis between physical basis and flight parameters, use multicollinearity analysis and stepwise regression method to select flight parameters with high correlation with center of gravity normal overload and center of gravity lateral overload. The selected flight parameters include: time, altitude, Mach, ascent and descent rate, normal overload, lateral overload, bank angle, pitch angle, heading angle, roll angular velocity, pitch angular velocity, and yaw angular velocity data;

[0033] Step 2: Time series flight parameter data processing.

[0034] S2.1: Data cleaning. Eliminate bit errors and out-of-bounds data for each takeoff and landing, perform small range filtering, and perform peak filtering. Perform air segment data discrimination on the flight parameter data based on the flight altitude, eliminate ground segment data, and only retain air segment flight parameter data to reduce the input data size.

[0035] S2.2: Data enhancement. Multiply the flight parameter data processed in step 2 by two parameters except the time term to obtain the product feature term between the flight parameter data;

[0036] S2.3: Data compression. Use multicollinearity diagnosis, amplitude filtering, and peak filtering to reduce the size of the sample space;

[0037] S2.4: Data standardization. Since the input data is not uniform in magnitude, such as the weight of the whole machine is in the order of 10,000, which leads to its natural weight, in order to minimize the magnitude of the input parameters, the data is standardized. At the same time, standardizing the data can eliminate the influence of noise. The data standardization method is shown in formula (1).

[0038]

[0039] Among them, x is the standardized data; μ is the mean value of the data; σ is the standard deviation of the data;

[0040] Step 3: Extract features from the time series flight parameter data processed in step 2.

[0041] S3.1: The characteristics of "time" are taken as take-off time, landing time, and total flight time; the characteristic values ​​of "full aircraft weight" are taken as take-off full aircraft weight and landing full aircraft weight; and the characteristic values ​​of "external weight" are taken as take-off external weight and landing external weight as time series data characteristic values;

[0042] S3.2: Except for the three parameters mentioned in S3.1, the maximum value, minimum value, average value and variance of the remaining parameters are extracted as the characteristic values ​​of their time series data;

[0043] S3.3: Using the normal overload at the center of gravity and the total weight of the aircraft as input, calculate the equivalent damage and nominal damage of the takeoff and landing, which are also used as the characteristic data of the takeoff and landing. Among them, the nominal damage calculation uses the normal overload at the center of gravity N in the flight parameters. z (nominal overload), equivalent damage adopts equivalent overload N z,当量 Calculation. The equivalent overload calculation is as shown in formula (2), where N z,名义 is the nominal overload, N z,当量 is the equivalent overload, W is N z,名义 The corresponding total machine weight, W 设计 Design weight for strength:

[0044] N z,当量 =N z,名义 ×WW 设计 (2)

[0045] The equivalent damage calculation method is as follows:

[0046] 1) For the selected load spectrum, the complete overload cycle is extracted by rain flow counting method.

[0047] The original overload spectrum is processed by rain flow counting and appears as an overload cycle. A combination of is the cycle characteristic (overload ratio) of the i-th overload cycle.

[0048] 2) The overload cycles at each level in the overload spectrum are converted into overload cycles under the pulsating cycle (R=0) according to the Odin transformation.

[0049] The Odin transformation formula is as follows:

[0050]

[0051] In formula (3), N zmax Indicates the peak value of stress cycle, R = N zmax / N zmin is the stress ratio, then the equivalent load corresponding to the i-th cycle is

[0052]

[0053] 3) Calculate the equivalent damage of each cycle based on the S-N curve (m=4) under the pulsating cycle;

[0054] Under the action of pulsating load cycles, the stress fatigue performance S-N curve of metal materials can be described in the form of a power function, expressed as:

[0055] S m N=C (5)

[0056] In formula (5), N represents the number of cycles and S represents the stress. When calculating nominal damage or equivalent damage, N is used.z,名义 or N z当量 Replace S and calculate the damage of each cycle. Then the damage corresponding to the i-th cycle can be expressed as:

[0057] D i =(N zmax ) 0.i m (6)

[0058] The total damage value is recorded as D n ,but

[0059]

[0060] S3.4: Combine the flight parameter characteristic values ​​of each take-off and landing sequence obtained in S3.2 and S3.3 with the corresponding flight subject information to construct a flight parameter characteristic & flight subject database;

[0061] Step 4: Use machine learning methods such as support vector machines and random forests to build a flight subject prediction model based on time series flight parameter data values ​​to achieve flight subject identification.

[0062] Example 1: Taking a certain type of aircraft as an example, the technical process of the specific implementation of the present invention is as follows Figure 1 As shown, the flight subject recognition model is constructed based on the design concept of the present invention. There are 6344 take-off and landing data of this type of aircraft, of which 2605 take-offs and landings are missing or obviously mislabeled flight subjects, which are eliminated, leaving 3739 take-offs and landings. Since it is impossible to identify the take-offs and landings that are not missing but mislabeled flight subjects, the flight subject recognition model is constructed based on the 3739 take-off and landing data with flight subject information:

[0063] 1) According to step 1, take the time series flight parameter data and flight subject information as input, Figure 2 The data processing method shown in the figure inputs the multivariate time series flight parameter data into the flight parameter screening. The 14 flight parameters screened are: time, altitude, Mach, ascent and descent rate, normal overload, lateral overload, bank angle, pitch angle, heading angle, roll angular velocity, pitch angular velocity, and yaw angular velocity data.

[0064] For the multivariate flight parameters, the stepwise regression method and the principal component analysis method were used to determine the flight parameters required for flight subject identification, which are: equivalent overload, roll rate voting signal, pitch rate voting signal, yaw rate voting signal, flight altitude, Mach number, rudder position signal, left outer elevon position signal, left inner elevon position signal, right outer elevon position signal, right inner elevon position signal, left leading edge flap position signal, right leading edge flap position signal, and lateral acceleration voting signal.

[0065] 2) According to step 2, firstly, except for "time", the flight parameter data are multiplied two by two to generate new parameter items, such as flight altitude × Mach number, Mach number × rudder position signal, rudder position signal × left outer elevator position signal and other nonlinear items; then, multicollinearity diagnosis, amplitude filtering and peak filtering are used to reduce the size of the sample space (dimensionality reduction) and standardize the data.

[0066] 3) According to step three, extract the features of the time-series flight parameter data of 3739 take-offs and landings. Among them, the features of "time" are taken as take-off time, landing time, and total flight time, the features of "weight" are taken as take-off weight and landing weight, and the features of the remaining flight parameters are taken as maximum value, minimum value, standard deviation, and mean value. The nominal damage is calculated using the nominal overload data, and the equivalent overload is calculated using the nominal overload combined with the weight, and the equivalent damage is calculated; and combined with the information of each take-off and landing flight subject, a "flight parameter feature data and flight subject database" is constructed, such as Figure 3 shown.

[0067] 4) According to step 4: 3739 take-off and landing data in the "flight parameter feature data and flight subject database" are randomly divided into 2991 take-off and landing (80%) feature data to construct a model for predicting flight subjects by inputting flight parameter features. The prediction model uses a support vector machine model, and the support vector machine selects a quadratic polynomial, a penalty coefficient C of 5.0, and polynomial independent terms coef0 = 5.0 to train the flight subject prediction model. After the model training is completed, the remaining 748 (20%) take-off and landing data that did not participate in the model training are used to test the model effect. Since the current sample data is not sufficient and there are human labeling errors in the samples, the accuracy of the constructed model is 93.5%.

[0068] The above is only one of the specific embodiments of the present invention, and the present invention is described in detail. The unexplained part is a conventional technology. However, the protection scope of the present invention is not limited to this. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. The protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for aircraft flight subject identification based on time series flight parameter data, characterized in that: The method comprises the following steps: Step S1: Screening and cleaning of flight parameter data; initially retaining time, external weight, weight, normal overload at the center of gravity and lateral overload at the center of gravity by default; performing correlation analysis between physical basis and flight parameters, using stepwise regression method for the remaining flight parameters, screening out flight parameters with high correlation with normal overload at the center of gravity and lateral overload at the center of gravity; then eliminating error codes, out-of-bounds data and ground segment data of the flight parameters to reduce the overall data size; Step S2: data processing; except for the time item, each flight parameter data processed in step S1 is used to generate new parameter items according to the physical meaning, mathematical and / or mechanical model of the flight parameters; then, multicollinearity diagnosis and filtering methods are used to reduce the size of the sample space, and the data is standardized to improve the stability of the model; Step S3: Extracting features from time series data; extracting features from the time series flight parameter data processed in step S2, and constructing a flight parameter feature database; taking the normal overload at the center of gravity and the total aircraft weight as input, calculating the equivalent damage and nominal damage of the takeoff and landing, which are also used as feature data of the takeoff and landing; and then combining the information of each takeoff and landing flight subject to construct a typical flight parameter feature & flight subject database; Step S4: Taking the typical flight parameter characteristics & flight subject database as input, a flight subject prediction model based on time series flight parameter data is constructed to realize flight subject identification.

2. The method for aircraft flight subject identification based on time series flight parameter data according to claim 1, characterized in that: When screening the flight parameter data in step 1, the flight parameter data is initially screened based on prior knowledge, and combined with the regression analysis algorithm to determine the final input flight parameter.

3. The method for aircraft flight subject identification based on time series flight parameter data according to claim 1, characterized in that: In step one, the error codes, out-of-bounds data removal, small range filtering, and peak filtering are performed on the flight parameters of each takeoff and landing. The flight parameter data is identified as the air segment by the flight altitude, the ground segment data is removed, and only the flight parameter data of the air segment is retained to reduce the input data scale.

4. The method for aircraft flight subject identification based on time series flight parameter data according to claim 1, characterized in that: In step 2, during data processing, the data is enhanced by the product term of the flying parameter data; and considering the possibility of strong linear correlation between the data after data enhancement, which will cause unnecessary increase in model training time, multicollinearity diagnosis is carried out to retain one group of multiple groups of data with strong linear correlation, thereby achieving dimensionality reduction of the model input independent variables.

5. The method for aircraft flight subject identification based on time series flight parameter data according to claim 4, characterized in that: The peak-valley filtering algorithm and the small-range filtering algorithm are used to filter the model input data based on the center of gravity overload, further reducing the scale of model training input data.

6. The method for aircraft flight subject identification based on time series flight parameter data according to claim 1, characterized in that: Feature extraction of time series flight parameter data, where the features of "time" are taken as take-off time, landing time and total flight time, the feature values ​​of "total aircraft weight" are taken as take-off weight and landing weight, and the feature values ​​of "external weight" are taken as take-off external weight and landing external weight; for the remaining flight parameters except time, weight and external weight, the features are taken as their maximum value, minimum value, standard deviation and mean.

7. The method for aircraft flight subject identification based on time series flight parameter data according to claim 1, characterized in that: In step three, the equivalent damage algorithm is used, and the "equivalent damage" and "nominal damage" calculated using the normal overload history of the center of gravity and the weight history of the entire aircraft are also used as the key features of the take-off and landing.

8. The method for aircraft flight subject identification based on time series flight parameter data according to claim 1, characterized in that: By using machine learning methods such as support vector machine and random forest, a flight subject prediction model based on the time series flight parameter feature value of machine learning models such as support vector machine and random forest is constructed.

9. The method for aircraft flight subject identification based on time series flight parameter data according to claim 1, characterized in that: In step S1, the 14 flight parameters finally screened include: time, external weight, weight, center of gravity normal overload, center of gravity lateral overload, altitude, Mach number, ascent and descent rate, roll angular velocity, pitch angular velocity, yaw angular velocity, bank angle, pitch angle, and heading angle.

Images