Helicopter fuselage vibration prediction method and device based on data driving

Abstract

The invention provides a helicopter fuselage vibration prediction method and device based on data driving, and the method comprises the following steps: 1, collecting test flight data of a helicopter test, the test flight data of the helicopter test comprising flight data and vibration data; 2, preprocessing the helicopter test flight test data, including flight / vibration data frequency collaboration, abnormal point elimination, five-point cubic smoothing and normalization, to obtain a preprocessed data set; step 3, performing variational mode decomposition on the vibration data to obtain each decomposed intrinsic mode function component, and calculating a weight coefficient of each I MF component by using a K-means algorithm; performing weighted reconstruction on the vibration data by using the weight coefficient to obtain a vibration data set after weighted reconstruction, and reducing the influence of noise components in the signal; and step 4, establishing a limit lifting decision tree model by using the preprocessed data set and the vibration data set after weighted reconstruction, and predicting helicopter fuselage vibration based on the limit lifting decision tree model.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent analysis of helicopter dynamics, and in particular relates to a method and device for predicting helicopter fuselage vibration based on data drive. Background Art

[0002] The vibration problem of the helicopter will increase the operating burden of the pilot and affect the flight quality. At the same time, excessive vibration will not only reduce the riding comfort, but also may cause fatigue damage to the structural components, seriously affecting the comfort and safety during the flight. With the development trend of high-speed, high-informatization and high-intelligence helicopters, the dynamic environment of the equipment is becoming more and more complex, and the vibration problem of the helicopter is also more complex and prominent. With the continuous improvement of safety and comfort requirements during flight, the accurate prediction of fuselage vibration has become an urgent need for the development of the industry.

[0003] There are two types of vibration analysis models for helicopter fuselage: one is the vibration mechanism analysis model based on the vibration mathematical model, and the other is the intelligent algorithm model based on data mining technology that has recently emerged. Due to the complex structure, strong coupling, and nonlinear characteristics of helicopters, the current helicopter vibration mechanism analysis model makes a lot of engineering assumptions during the modeling process, resulting in a certain error between the vibration prediction results of the vibration mechanism analysis model and the actual measured values. At the same time, the operation time of the mechanism model is relatively long and cannot provide real-time warning. Compared with the traditional mechanism model, the prediction results of the data-driven intelligent analysis method are more accurate and easier to realize the control and diagnosis of complex equipment. However, there are few application examples of intelligent algorithms in helicopter vibration research. Summary of the invention

[0004] In response to the above technical problems, the present invention provides a data-driven method and device for predicting helicopter fuselage vibration, which overcomes the defects of the traditional mechanism model, such as the complex modeling process, the difficulty of vibration prediction and the poor accuracy, and can meet the needs of short-term and high-precision prediction of the helicopter fuselage.

[0005] In a first aspect, the present application provides a data-driven helicopter fuselage vibration prediction method, the method comprising the following steps:

[0006] Step 1: Collect helicopter test flight data, wherein the helicopter test flight data includes flight data and vibration data;

[0007] Step 2: Preprocess the helicopter test flight data, including flight / vibration data frequency coordination, outlier removal, five-point cubic smoothing and normalization, to obtain a preprocessed data set;

[0008] Step 3: Perform variational mode decomposition (VMD) on the vibration data to obtain the decomposed intrinsic mode functions (IMF) components, and use the K-means algorithm to calculate the weight coefficient of each IMF component; use the weight coefficient to weighted reconstruct the vibration data to obtain a weighted reconstructed vibration data set to reduce the influence of noise components in the signal;

[0009] Step 4: Using the preprocessed data set and the weighted reconstructed vibration data set, an extreme gradient boosting regression (Xgboost) model is established, and the helicopter fuselage vibration is predicted based on the extreme gradient boosting regression model.

[0010] Furthermore, the specific method of step 2 is:

[0011] Step 2.1: Perform frequency coordination on the sampling frequencies of different types of data, use the target frequency to downsample the vibration data, and perform cubic spline interpolation expansion on the flight data to achieve frequency coordination between different samples;

[0012] Step 2.2 uses the Pauta criterion to remove outliers in the helicopter test flight data:

[0013]

[0014]

[0015] Where: y i is the experimental measurement value of the ith sample, is the mean value of a specific variable, and n is the total number of samples;

[0016] Step 2.3: Use the five-point cubic smoothing method to remove high-frequency vibration noise in the helicopter test flight data to improve the overall quality of the sample data:

[0017]

[0018] Where: is the value after data smoothing;

[0019] Step 2.4: Perform Z-score standardization on the helicopter test flight data. The formula is as follows:

[0020]

[0021] Where: y ii is the normalized value of the ith sample, μ xis the mean of the sample data after smoothing.

[0022] Furthermore, the specific method of step 3 is:

[0023] Step 3.1: Perform fast Fourier transformation (FFT) on the preprocessed vibration data to obtain the main frequency characteristics of the vibration data;

[0024] Step 3.2: For the vibration data, use variational mode decomposition (VMD) to decompose the vibration data into K different intrinsic mode functions (IMFs), as shown in the following formula:

[0025] y k (t) = A k (t)cos[φ k (t)] (5)

[0026] Where: A k (t) is the amplitude of the kth IMF component, φ k (t) is the phase of the kth IMF component;

[0027] Among them, the calculation method of the number of VMD decomposition modes K is:

[0028] Traverse and search K∈[2,3,4,5,6,7,8,9]. As K increases, when the decomposed IMF component first appears consistent with the main frequency characteristics in the vibration data in step 3.1, the optimal decomposition mode number K is obtained;

[0029] Step 3.3: Based on K different intrinsic mode functions (IMFs), plot the distribution of the mean and variance of each IMF component. Based on the distribution of the mean and variance of each IMF signal, use the Euclidean distance formula and the K-means algorithm to calculate the weight of each IMF component. The specific formula is as follows:

[0030]

[0031] Where: is the average value of the ith IMF component after decomposition, v i is the variance of the i-th IMF component after decomposition;

[0032] Step 3.4: Use the weights of each IMF component to complete the weighted reconstruction of the vibration data. The specific formula is as follows:

[0033] y(t)=ω 1 y 1 (t)+ω 2 y 2 (t)+…+ω ky k (t) (7)

[0034] Where: y(t) is the weighted reconstructed vibration data, ω 1 is the weight of the first IMF component, y 1 (t) is the first IMF component after decomposition.

[0035] Furthermore, the specific method of step 4 is:

[0036] Step 4.1: Combine the preprocessed data set and the weighted reconstructed vibration data set to obtain a data sample set for intelligent algorithm prediction;

[0037] Step 4.2: Adopt the adaptive cuckoo search algorithm (CSA) to determine the number of loops n and the maximum decision tree depth m in the Xgboost model, traverse the search m∈[10:10:500] and n∈[2:1:10] to determine the optimal number of loops and maximum decision tree depth;

[0038] Step 4.3: Use the CSA method to determine the L1 and L2 regularization coefficients in the Xgboost model, traverse and search L1∈[0:0.1:1] and L2∈[0:0.1:1], and determine the optimal number of cycles and the maximum decision tree depth;

[0039] Step 4.4: Use the CSA method to determine the learning rate α in the Xgboost model, traverse and search α∈[0.1:0.1:1], and determine the optimal learning rate;

[0040] Step 4.5: Based on the Xgboost model, predict the vibration acceleration of the fuselage.

[0041] Furthermore, the flight data include total pitch displacement, longitudinal operating displacement, lateral operating displacement, pedal displacement, heading angle, pitch angle, roll angle, true airspeed, absolute pressure altitude, main rotor shaft torque, rotor total pitch, horizontal tail deflection angle, tail rotor pitch, constant atmospheric temperature, constant atmospheric pressure, indicated airspeed, vertical speed, longitudinal overload at the center of gravity, lateral overload at the center of gravity, and normal overload at the center of gravity.

[0042] Furthermore, the vibration data includes vibration acceleration at the floor of the right driver's seat.

[0043] Furthermore, step 4.1 specifically includes:

[0044] 75% of the data in the data sample set are randomly used as training samples to obtain the network structure of the Xgboost model, and the remaining 25% of the samples in the data sample set are used as test sets.

[0045] In a second aspect, the present application provides a data-driven helicopter fuselage vibration prediction device, the device comprising a collection unit, a preprocessing unit, a weighted reconstruction unit and a prediction unit, wherein:

[0046] A collection unit, used for collecting helicopter test flight data, wherein the helicopter test flight data includes flight data and vibration data;

[0047] A preprocessing unit is used to preprocess the helicopter test flight data, including flight / vibration data frequency coordination, outlier removal, five-point cubic smoothing and normalization, to obtain a preprocessed data set;

[0048] The weighted reconstruction unit is used to perform variational mode decomposition (VMD) on the vibration data to obtain each intrinsic mode function (IMF) component after decomposition, and calculate the weight coefficient of each IMF component using the K-means algorithm; the vibration data is weighted reconstructed using the weight coefficient to obtain a weighted reconstructed vibration data set to reduce the influence of noise components in the signal;

[0049] The prediction unit is used to establish an extreme gradient boosting regression (Xgboost) model by using the preprocessed data set and the weighted reconstructed vibration data set, and predict the vibration of the helicopter fuselage based on the extreme gradient boosting regression model.

[0050] In summary, the present invention proposes a data-driven helicopter fuselage vibration prediction method and device, which effectively solves the problems of multi-source heterogeneous data sampling frequency coordination, abnormal points and order of magnitude differences through comprehensive preprocessing of flight / vibration data; adopts weighted VMD decomposition and reconstruction technology to reduce noise interference in vibration data; based on the data set after comprehensive preprocessing and VMD decomposition and weighted reconstruction, uses the Xgboost model and combines the adaptive cuckoo search method to design a high-precision helicopter fuselage vibration intelligent prediction algorithm, which overcomes the defects of complex modeling process and poor vibration prediction accuracy of traditional mechanism models, and can meet the short-term high-precision prediction needs of helicopter fuselage. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flow chart provided by an embodiment of the present invention;

[0052] Figure 2 is the center frequency of the decomposed IMF component under different K values;

[0053] Figure 3 It is the distribution diagram of mean and variance of IMF components after decomposition. DETAILED DESCRIPTION

[0054] In order to achieve high-precision intelligent prediction of helicopter fuselage vibration response, the present invention comprehensively considers the influence of multiple flight variables on vibration, designs a high-precision and strong generalization ability intelligent prediction method for helicopter fuselage vibration, uses helicopter test flight data to establish a data-driven model to achieve accurate prediction of fuselage vibration response, and studies how to use intelligent algorithm models to realize helicopter vibration problem analysis.

[0055] like Figure 1 As shown, the present invention proposes a data-driven helicopter fuselage vibration prediction method for high-precision prediction of helicopter fuselage vibration response, and the specific steps include the following steps:

[0056] Step 1: using a collection unit to collect helicopter test flight data, the helicopter test flight data includes flight data and vibration data; selecting 35 flight variables that have the greatest impact on helicopter fuselage vibration as flight data, including total displacement, longitudinal operation displacement, lateral operation displacement, pedal displacement, heading angle, pitch angle, roll angle, true airspeed, absolute pressure altitude, main rotor shaft torque, rotor total pitch, horizontal tail deflection angle, tail rotor pitch, atmospheric constant temperature, atmospheric constant pressure, indicated airspeed, lifting speed, longitudinal overload at the center of gravity, lateral overload at the center of gravity, normal overload at the center of gravity, etc.; selecting the vibration acceleration data at the floor of the right driver's seat of the fuselage as vibration data;

[0057] Step 2: Preprocess the helicopter test flight data. The specific steps are as follows;

[0058] Step 2.1: Downsample the vibration data with a target frequency of 200 Hz and perform cubic spline interpolation expansion on the flight data to achieve frequency coordination between different samples;

[0059] Step 2.2: Use the Pauta criterion to remove outliers in the vibration data;

[0060]

[0061]

[0062] Where: y i is the experimental measurement value of the ith sample, is the mean value of a specific variable, and n is the total number of samples;

[0063] Step 2.3: Use the five-point cubic smoothing method to remove high-frequency vibration noise in the sample data;

[0064]

[0065] Where: is the value after data smoothing;

[0066] Step 2.4: Perform Z-score normalization on the flight and vibration data. The formula is as follows:

[0067]

[0068] Where: y ii is the normalized value of the ith sample, μ x is the mean of the sample data after smoothing.

[0069] Step 3: Use variational mode decomposition (VMD) to reconstruct the vibration acceleration data in a weighted manner. The specific steps are as follows:

[0070] Step 3.1: Perform fast Fourier transformation (FFT) on the vibration data preprocessed in step 2 to obtain the main frequency characteristics of the vibration data.

[0071] Step 3.2: For the vibration data, first use VMD decomposition to decompose the original vibration data into K different intrinsic mode functions (IMFs), the formula is as follows:

[0072] y k (t) = A k (t)cos[φ k (t)] (5)

[0073] Where: A k (t) is the amplitude of the kth IMF component, φ k (t) is the phase of the kth IMF component;

[0074] Among them, the calculation method of the number of VMD decomposition modes K is:

[0075] Traverse and search K∈[2,3,4,5,6,7,8,9]. As K increases, when the decomposed IMF component first appears consistent with the main frequency characteristics in the vibration data in step 3.1, the optimal decomposition mode number K is obtained.

[0076] Determine the value of the VMD decomposition mode number K, and traverse and search K∈[2,3,4,5,6,7,8,9], Figure 2is the center frequency of the IMF component after decomposition under different K values. It can be found that when K = 6, the principal components of 3 and 5Ω can be effectively separated from the original vibration data. Therefore, the VMD decomposition mode number K value is 6, and the original vibration data is decomposed into 6 different IMF components.

[0077] Step 3.4: According to the K IMF components obtained after VMD decomposition, draw the distribution of the mean and variance of each IMF component. According to the distribution of the mean and variance of each IMF signal, use the Euclidean distance formula and the K-means algorithm to calculate the weight of each IMF component. The specific formula is as follows:

[0078]

[0079] Where: is the average value of the ith IMF component after decomposition, v i is the variance of the i-th IIMF component after decomposition.

[0080] according to Figure 3 According to the distribution of the mean and variance of each IMF component in the decomposed IMF signal, the decomposed IMF signal can be divided into three categories: the first category is the 3rd to 6th IMF components, the second category is the 1st IMF component, and the third category is the 2nd IMF component. For the second and third categories of IMF components after division, their weights are 1 in the signal weighted reconstruction process. For the first category of signals, the 6th IMF component with a larger mean and a smaller variance is selected as the weight calculation standard, and its weight is 1. According to formula (6), the Euclidean distance D between the 3rd to 5th IMF and IMF 6 is calculated respectively to obtain the weights of the 3rd to 5th IMF. The final weights of IMF 1-6 are:

[0081] ω 1 =1,ω 2 =1,ω 3 =0.77,ω 4 =0.89,ω 5 =0.95,ω 6 =1 (7)

[0082] Step 3.5: Use the weight calculation results of each IMF component to complete the weighted reconstruction of the vibration data. The specific formula is as follows:

[0083]

[0084] Where: y(t) is the weighted reconstructed vibration data, ω 1 is the weight of the first IMF component, y 1 (t) is the first IMF component after decomposition.

[0085] 4. The method according to claim 1, characterized in that: the specific method of step 4 is:

[0086] Step 4.1: Combine the preprocessed data set and the weighted reconstructed vibration data set to obtain a data sample set for intelligent algorithm prediction.

[0087] 75% of the data in the sample set are randomly used as training samples to obtain the best network structure parameters of the Xgboost model, and the remaining 25% of the data are used as the prediction set.

[0088] Step 4.2: Adopt the adaptive cuckoo search algorithm (CSA) to determine the number of loops n and the maximum decision tree depth m in the Xgboost model, traverse the search m∈[10:10:500] and n∈[2:1:10], and determine the optimal number of loops and maximum decision tree depth.

[0089] When m=3 and n=7, the prediction result determination coefficient R2 of the Xgboost model has the highest accuracy. Therefore, the optimal number of cycles determined by the CSA method is 70 and the maximum decision tree depth is 3.

[0090] Step 4.3: Use the CSA method to determine the L1 and L2 regularization coefficients in the Xgboost model, traverse and search L1∈[0:0.1:1] and L2∈[0:0.1:1], and determine the optimal number of cycles and the maximum decision tree depth.

[0091] When L1=0 and L2=1, the prediction result determination coefficient R2 of the Xgboost model has the highest accuracy. Therefore, the optimal L1 regularization coefficient determined by the CSA method is 0 and the L2 regularization coefficient is 1.

[0092] Step 4.4: Use the CSA method to determine the learning rate α in the Xgboost model, traverse the search α∈[0.1:0.1:1], and determine the optimal learning rate.

[0093] The CSA method is used to determine the learning rate α in the Xgboost model. With an increment of 0.1, traverse the search for α∈[0.1,0.1,1], and record the prediction result determination coefficient R2 in 10 calculations. When α=0.3, the prediction result determination coefficient R2 of the Xgboost model has the highest accuracy. Therefore, the optimal learning rate α determined by the CSA method is 0.3.

[0094] Step 4.5: Based on the Xgboost model, the fuselage vibration acceleration is predicted using the data sample set obtained in step 4.1.

[0095] The average percentage absolute error of the prediction results of the vibration acceleration at the right driver's seat floor by the method proposed in the present invention is 4.49%, and the relative error of the prediction results of 92.9% of the sample points is less than ±10%.

Claims

1. A data-driven helicopter fuselage vibration prediction method, characterized in that: The method comprises the following steps: Step 1: Collect helicopter test flight data, wherein the helicopter test flight data includes flight data and vibration data; Step 2: Preprocess the helicopter test flight data, including flight / vibration data frequency coordination, outlier removal, five-point cubic smoothing and normalization, to obtain a preprocessed data set; Step 3: Perform variational mode decomposition on the vibration data to obtain the intrinsic mode function components after decomposition, and use the K-means algorithm to calculate the weight coefficient of each IMF component; use the weight coefficient to perform weighted reconstruction on the vibration data to obtain a weighted reconstructed vibration data set to reduce the impact of noise components in the signal; Step 4: Using the preprocessed data set and the weighted reconstructed vibration data set, an extreme lifting decision tree model is established, and the helicopter fuselage vibration is predicted based on the extreme lifting decision tree model.

2. The method according to claim 1, characterized in that The specific method of step 2 is: Step 2.1: Perform frequency coordination on the sampling frequencies of different types of data, use the target frequency to downsample the vibration data, and perform cubic spline interpolation expansion on the flight data to achieve frequency coordination between different samples; Step 2.2 uses the Pauta criterion to remove outliers in the helicopter test flight data: Where: y i is the experimental measurement value of the ith sample, is the mean value of a specific variable, and n is the total number of samples; Step 2.3: Use the five-point cubic smoothing method to remove high-frequency vibration noise in the helicopter test flight data to improve the overall quality of the sample data: Where: is the value after data smoothing; Step 2.4: Perform Z-score standardization on the helicopter test flight data. The formula is as follows: Where: y ii is the normalized value of the ith sample, μ x is the mean of the sample data after smoothing.

3. The method according to claim 1, characterized in that , the specific method of step 3 is: Step 3.1: Perform fast Fourier transform on the preprocessed vibration data to obtain the main frequency characteristics of the vibration data; Step 3.2: For the vibration data, use variational mode decomposition to decompose the vibration data into K different eigenmode functions. The formula is as follows: y k (t)=A k (t)cos[φ k (t)] (5) Where: A k (t) is the amplitude of the kth IMF component, φ k (t) is the phase of the kth IMF component; Among them, the calculation method of the number of VMD decomposition modes K is: Traverse and search K∈[2,3,4,5,6,7,8,9]. As K increases, when the decomposed IMF component first appears consistent with the main frequency characteristics in the vibration data in step 3.1, the optimal decomposition mode number K is obtained; Step 3.3: According to K different intrinsic mode functions, draw the distribution of the mean and variance of each IMF component. According to the distribution of the mean and variance of each IMF signal, use the Euclidean distance formula and K-means algorithm to calculate the weight of each IMF component. The specific formula is as follows: Where: is the average value of the ith IMF component after decomposition, v i is the variance of the i-th IMF component after decomposition; Step 3.4: Use the weights of each IMF component to complete the weighted reconstruction of the vibration data. The specific formula is as follows: y(t)=ω1y1(t)+ω2y2(t)+…+ω k y k (t) (7) Where y(t) is the weighted reconstructed vibration data, ω1 is the weight of the first IMF component, and y1(t) is the first IMF component after decomposition.

4. The method according to claim 1, characterized in that: The specific method of step 4 is: Step 4.1: Combine the preprocessed data set and the weighted reconstructed vibration data set to obtain a data sample set for intelligent algorithm prediction; Step 4.2: Adopt the adaptive cuckoo search method to determine the number of loops n and the maximum decision tree depth m in the Xgboost model, traverse the search m∈[10:10:500] and n∈[2:1:10], and determine the optimal number of loops and maximum decision tree depth; Step 4.3: Use the CSA method to determine the L1 and L2 regularization coefficients in the Xgboost model, traverse and search L1∈[0:0.1:1] and L2∈[0:0.1:1], and determine the optimal number of cycles and the maximum decision tree depth; Step 4.4: Use the CSA method to determine the learning rate α in the Xgboost model, traverse and search α∈[0.1:0.1:1] to determine the optimal learning rate; Step 4.5: Based on the Xgboost model, predict the vibration acceleration of the fuselage.

5. The method according to claim 1, characterized in that: The flight data includes total displacement, longitudinal operating displacement, lateral operating displacement, pedal displacement, heading angle, pitch angle, roll angle, true airspeed, absolute pressure altitude, main rotor shaft torque, total rotor pitch, horizontal tail deflection angle, tail rotor pitch, constant atmospheric temperature, constant atmospheric pressure, indicated airspeed, vertical speed, longitudinal overload at the center of gravity, lateral overload at the center of gravity, and normal overload at the center of gravity.

6. The method according to claim 1, characterized in that The vibration data includes the vibration acceleration at the right driver's seat floor.

7. The method according to claim 1, characterized in that: Step 4.1 specifically includes: 75% of the data in the data sample set are randomly used as training samples to obtain the network structure of the Xgboost model, and the remaining 25% of the samples in the data sample set are used as test sets.

8. A data-driven helicopter fuselage vibration prediction device, characterized in that: The device comprises a collection unit, a preprocessing unit, a weighted reconstruction unit and a prediction unit, wherein: A collection unit, used for collecting helicopter test flight data, wherein the helicopter test flight data includes flight data and vibration data; A preprocessing unit is used to preprocess the helicopter test flight data, including flight / vibration data frequency coordination, outlier removal, five-point cubic smoothing and normalization, to obtain a preprocessed data set; The weighted reconstruction unit is used to perform variational mode decomposition (VMD) on the vibration data to obtain each intrinsic mode function (IMF) component after decomposition, and calculate the weight coefficient of each IMF component using the K-means algorithm; the vibration data is weighted reconstructed using the weight coefficient to obtain a weighted reconstructed vibration data set to reduce the influence of noise components in the signal; The prediction unit is used to establish an extreme gradient boosting regression (Xgboost) model by using the preprocessed data set and the weighted reconstructed vibration data set, and predict the vibration of the helicopter fuselage based on the extreme gradient boosting regression model.

Images