CN103335814A - Inclination angle measurement error data correction system and method of experimental model in wind tunnel - Google Patents

Abstract

The invention discloses an inclination angle measurement error data correction system and method of an experimental model in a wind tunnel when the experimental model is placed in the wind tunnel to be subjected to a blowing experiment. The inclination angle measurement data correction system comprises a data preprocessing module, a model training module and an error data correction module, and the correction method comprises three steps: data preprocessing, model training and error data correcting. The invention provides a multi-stage and multi-model fusion error correction method based on a neural network model as a basic model, the model fitting degree is higher, the calculation result is more accurate, the nonlinear error, caused by vibration of the model under the wind action, in data measured actually by an inclination angle sensor can be corrected well, and the correction error of the correction system can be controlled within 0.004-0.01 degree near the 0 degree.

Description

System and method for correcting error data of inclination angle measurement of experimental model in wind tunnel

technical field

The invention relates to an experimental model in a wind tunnel, a correction system and a correction method for the inclination angle measurement error data of the experimental model in the wind tunnel during the blowing experiment process.

Background technique

At present, the measurement technology of supersonic and transonic wind tunnel models is more and more used in the production and testing of civilian products, and the requirements and accuracy of testing technology are getting higher and higher. Since the wind tunnel test process is a complex aerodynamic experiment with many links, it is affected by many factors such as test tools, test environment, and tester's technical mastery level, which will affect the test conclusion, test precision and accuracy. and error. How to obtain accurate test data and eliminate the influence of the environment on accuracy is crucial to the measurement technology of supersonic and transonic wind tunnels.

In the wind tunnel model, there is an angle α between the bracket and the horizontal plane, which is called the inclination angle, which can be measured by the gravity sensor inside the model. The inclination sensor is used to measure the actual angle of attack of the model after loading, and is located at the forefront of the detection system. The measured value of the angle of attack is one of the basic data for other numerical operations of the detection system, and its accuracy directly affects the accuracy of the test data of the detection system. In the actual wind tunnel test, the model will be affected by external forces from the front, rear, left and right, and the torque generated by the external force will act on the model, which will cause errors in the measurement of the gravity sensor of the model's inclination angle. That is, under the influence of various external force factors, there are errors in the data measured by the inclination sensor. After analysis, this error is mainly due to the fact that the model is subjected to vibration moments in different directions, so that the working environment of the inclination sensor is no longer stable, so there is an error in the digital signal obtained by the sensor, which cannot truly reflect the angle of attack of the model value. Therefore, the original digital signal obtained by the inclination sensor cannot be directly used for other numerical calculations, and this signal needs to be corrected and compensated to ensure that the test data can truly reflect the tested situation.

In order to reduce the amount of error, nonlinear mathematical models are often used when modeling compensation systems, and linear models are generally not used. In the nonlinear model system, the commonly used data error compensation methods include successive approximation method, least squares polynomial curve fitting method, adaptive filtering method, fitting error compensation formula method, cubic spline interpolation and neural network method, etc. .

Due to the complexity and inevitability of the torque generated by the vibration of the model by the wind, the error of the inclination sensor data due to the vibration of the model always exists, and this error is usually a complex function or model of multiple variables. If the conventional mathematical fitting method is used to obtain this function, then the key problem to be solved is not only to make a reasonable mathematical description of the possible variables, but also to explore as many variables as possible in the environment. It is very difficult to define the exact mathematical model of the vibration error in the wind tunnel to describe the relationship between these variables, and the expansion ability of the model cannot be guaranteed. Recently, in the field of data error correction technology, a new technology of using neural network technology for error compensation value fitting has been proposed. Since the neural network does not require an accurate mathematical model, it has good nonlinear mapping ability, and is good at learning useful knowledge from input and output data, revealing data characteristics, and dealing with random factors. Using the neural network method to solve how to describe the relationship between the data environment and the data error, through the sufficient training of the neural network, using the ability of the neural network to self-organize and optimize the modeling, and continuously carry out self-organization optimization with the training data, thus forming The optimal and adaptable error compensation model improves the ability to correct accurate error data.

Contents of the invention

The purpose of the present invention is to improve the performance of the existing error compensation technology, to provide a method that can effectively correct the non-linear error caused by vibration in the measured data of the inclination sensor, and the correction error of the inclination value can be controlled within 0.004 to 0.01 accuracy near the 0 angle. The invention relates to a correction system and correction method for the inclination angle measurement error data of the experimental model in the wind tunnel.

The object of the present invention is achieved through the following technical solutions: the experimental model inclination angle measurement error data correction system in the wind tunnel, which includes the following modules:

Data preprocessing module: Preprocessing the input data, extracting the vibration error feature vector, the system error correction model training process and the data correction process both include the data preprocessing module, but the preprocessing module in the error correction model training process and the data correction process complete different tasks;

Model training module: according to the training data and target data collected in advance, establish a joint error correction model with angle, frequency and amplitude as input variable parameters, and error correction value as output. This module includes model feature vector extraction sub-module and model training sub-module module;

Data correction module: According to the input data of each segment obtained by the data preprocessing module in the error correction process, the feature vector of each segment of input data is extracted, and the feature vector is input into the joint error correction model obtained in advance to obtain the error correction value, and the correction Values are added to the original input to obtain the corrected data.

The correction method of the inclination measurement error data of the experimental model in the wind tunnel includes three steps: data preprocessing, model training and data correction. Both model training and data correction include data preprocessing modules, but preprocessing in model training and data correction Modules perform different tasks;

Data preprocessing in the described model training step includes the following sub-steps:

S101: Load target data and training data to the correction system, and down-sample the target data curve and training data curve;

S102: respectively performing wavelet transform and smoothing processing on the target data curve and the training data curve to obtain their high-frequency and low-frequency data;

S103: Calculate the difference between the high and low frequency data of the target data curve and the high and low frequencies of all the training data curves to obtain the input (difference of high frequency) and output data set (difference of low frequency) used for training the model;

Data preprocessing in the described data correction step includes the following sub-steps:

S201: Load the target data and each data to be corrected to the system, and down-sample the target curve and the curve to be corrected;

S202: performing wavelet transformation and smoothing processing on the target curve and the curve to be corrected respectively, to obtain their high-frequency and low-frequency data;

S203: Calculate the difference between the high and low frequency data of the target curve and the high and low frequency of the curve to be corrected, and obtain the input (high frequency difference) data and the reference curve of the error correction model;

S204: Segment the obtained input data according to the result of the endpoint detection;

The described model training step comprises the following sub-steps:

S301: Extract corrected features according to the input data obtained by preprocessing in the model training step;

S302: The output data obtained by the preprocessing module in the model training process is used as the output value, and the corrected features obtained in S301 are used as input values to form a training data set. The system takes each dimension of the corrected feature vector value as a unit variable, and trains The training data subset representing the data change of each unit variable in the data set is divided into the model training set of the unit variable, and each unit variable is used as the input variable parameter for training, and the output data is N of the output value (N=unit variable number number) unit variation model and stored as an intermediate result;

S303: Using the center point of each eigenvalue in each unit variable training data subset obtained by dividing in S302, interpolate it within a certain range, and respectively obtain a new eigenvalue set of each unit variable after interpolation;

S304: Bring the new eigenvalue sets of each unit variable obtained in S303 into the unit change model obtained in S302 to obtain the output values of each model, and combine the new eigenvalue sets of each unit variable and the output values of each model together to form One-dimensionally change the training data set and store it for subsequent model training;

）个空间曲面，曲面中每个点为由2维单元变量值和对应模型输出值构成的一向量； S305: Using the one-dimensional change training data set obtained in S304, based on the Krging surface fitting algorithm, respectively fitting M with any two unit variable values and corresponding model output values as three-dimensional coordinates (M=

) a spatial surface, each point in the surface is a vector composed of 2-dimensional unit variable values and corresponding model output values;

S306: Sampling K points on each of the M spatial surfaces generated in S305 to form an error data training set of KxM elements, each point in the set is a four-dimensional vector, and use the error data training set to train Error correction model Q, which is stored in the system for error data correction process;

The data correction includes the following sub-steps:

S401: Extracting input data obtained by data preprocessing in the data correction step, and segmenting each input, and organizing the data into a matrix with a certain column length;

S402: Using the same method as S301 to extract the eigenvalues of each segment of data, and bringing the eigenvalues of each segment of data into the trained error correction model to obtain the error correction value of each segment of data;

S403: Connect the error correction values of each segment of data to form a complete error correction value, and perform upsampling to form the final error correction value;

S404: Add the final error correction amount to the original data with correction to obtain corrected inclination sensor data.

The unit variables in the model training step include frequency h, amplitude f and angle a.

The advantages of the present invention are:

1. Based on the neural network model to build a multi-stage, multi-model fusion error correction method, the constructed model has fast learning and can approach any internal model and internal model controller, so that the error correction of the inclination sensor data is very good, so that the model simulates The degree of fit is higher, and the result of error compensation is more accurate.

2. By training the neural network correction model under various vibration frequencies and vibration intensities, when the inclination sensor data is input, the error inclination sensor data is corrected based on the multi-model and multi-stage data correction algorithm; it can effectively correct the vibration caused by the actual measurement data of the inclination sensor The non-linear error caused by it, and the correction value of the inclination error can be controlled within the accuracy of 0.004-0.01 around the 0 angle.

Description of drawings

Fig. 1 is a system frame diagram of the present invention;

Figure 2 Flowchart of data preprocessing during system model training;

Figure 3 is a flow chart of data preprocessing in the process of system data correction;

Fig. 4 is a flow chart of the model training method of the present invention;

Fig. 5 is a schematic diagram of the neural network structure of the error data correction model;

Fig. 6 is a flow chart of data correction in the present invention.

Detailed ways

个倾角传感器数据。 Training data set: Vibrate the experimental model under the conditions of different frequencies, amplitudes and inclinations through the shock device in advance, and collect the data obtained at different frequencies, different amplitudes and different inclinations.

tilt sensor data.

Target data: artificially determined error-free inclination sensor data.

The technical solution of the present invention will be further described below in conjunction with the accompanying drawings, but the content protected by the present invention is not limited to the following description.

As shown in Figure 1, the inclination measurement error data correction system of the experimental model in the wind tunnel includes the following modules:

Data preprocessing module: Preprocessing the input data, extracting the vibration error feature vector, the system error correction model training process and the data correction process both include the data preprocessing module, but the preprocessing module in the error correction model training process and the data correction process complete different tasks;

Model training module: according to the training data and target data collected in advance, establish a joint error correction model with angle, frequency and amplitude as input variable parameters, and error correction value as output. This module includes model feature vector extraction sub-module and model training sub-module module;

Data correction module: According to the input data of each segment obtained by the data preprocessing module in the error correction process, the feature vector of each segment of input data is extracted, and the feature vector is input into the joint error correction model obtained in advance to obtain the error correction value, and the correction Values are added to the original input to obtain the corrected data.

The correction method of the inclination measurement error data of the experimental model in the wind tunnel includes three steps: data preprocessing, model training and data correction. Both model training and data correction include data preprocessing modules, but preprocessing in model training and data correction Modules perform different tasks;

As shown in Figure 2, the data preprocessing in the described model training step includes the following sub-steps:

倍下采样，获得下采样后的目标数据和各个训练数据的压缩数据； S101: Load the target data and training data to the correction system, and carry out the target data curve and the training data curve

times downsampling to obtain the downsampled target data and the compressed data of each training data;

S102: Perform continuous wavelet transform on each compressed data, and then perform smoothing and filtering processing to obtain two parts of frequency data of high frequency and low frequency of each compressed data;

（高频之差）和

（低频之差）（

 ）输出数据集合； S103: Calculate the difference between the high and low frequency data of the target data curve and the high and low frequencies of all training data curves to obtain the input for training the model

(difference in high frequency) and

(difference in low frequency) (

) output data set;

As shown in Figure 3, the data preprocessing in the data correction step includes the following sub-steps:

倍下采样，获得下采样后的目标数据和待修正数据的压缩数据；  S201: Load the target data and each data to be corrected to the system, and carry out the target curve and the curve to be corrected

times downsampling to obtain the downsampled target data and the compressed data of the data to be corrected;

S202: Perform continuous wavelet transform on the compressed data of the target data and the data to be corrected respectively, and then perform smoothing and filtering processing to obtain their high-frequency and low-frequency data;

S203: Calculate the difference between the high and low frequency curves of the target data and the high and low frequencies of the data to be corrected to obtain the input of the error correction model (difference in high frequency) data and reference curves.

分为

 段，并对每段后续分别进行补偿； S204: According to the result of endpoint detection, the

Divided into

segment, and compensate each subsequent segment separately;

As shown in Figure 4, the correction model training step includes the following sub-steps:

（

 ），提取修正特征，具体步骤如下： S301: Obtain according to the input data obtained by preprocessing in the model training step

(

), extract the corrected features, the specific steps are as follows:

进行短时傅里叶变换，获得

每个点的频率数据； S3011: yes

Perform short-time Fourier transform to obtain

Frequency data for each point;

={，

 ，

 }，

 ，P=的点数。其中

为对应的

频率数据的第

 点数值，

为对应的

幅度数据的第点数值，

为对应的第

 个训练数据的倾角值； S3012: Feature extraction: constructing P 3-dimensional feature vectors

={ ,

,

},

, P= points. in

for the corresponding

of the frequency data

pip value,

for the corresponding

the amplitude data pip value,

for the corresponding

The inclination value of the training data;

（

 ）为输出数据，S301中获得的修正特征

（

 ）为输入值，构成训练数据集合，系统以修正特征向量值

的每一维作为一单元变量，将训练数据集合中表征每一个单元变量数据变化的训练数据子集划分为该单元变量的模型训练集合，即

 变化训练数据子集{

，

|属于变化}、

 变化训练数据子集{

，|属于

变化}和

 变化训练数据子集{

，

|属于

变化}，分别训练以

为输入变量参数，输出数据为输出值 的3个单元变化模型

 ，并存储为中间结果； S302: Obtained by preprocessing in the model training step

(

) is the output data, the corrected features obtained in S301

(

) as the input value to form a training data set, and the system corrects the eigenvector value

Each dimension of is used as a unit variable, and the training data subset representing the data change of each unit variable in the training data set is divided into the model training set of the unit variable, namely

change training_data_subset {

,

| belongs to Variety},

change training_data_subset {

, | belongs to

change} and

change training_data_subset {

,

| belongs to

Variation}, respectively trained with

is the input variable parameter, and the output data is the output value The 3 unit change model of

, and stored as an intermediate result;

S303: Use the center point of each eigenvalue in each unit variable training data subset obtained by dividing in S302, use the following formula to interpolate it within a certain range, and obtain new eigenvalue sets of each unit variable after interpolation respectively:

=  

 

=

=

=

=

，并将新的各个单元变量特征值集合和各个模型输出值，一同构成一维变化训练数据集合{

，

}，并存储用于后续的模型训练； S304: Collect the new eigenvalues of each unit variable obtained in S303 respectively into the unit change model obtained in S302 Get the output value of each model

, and combine the new feature value sets of each unit variable and each model output value together to form a one-dimensional change training data set{

,

}, and stored for subsequent model training;

， ）、（

，

，

 ）（

，

，

）3个空间曲面，曲面中每个点为由2维单元变量值和对应模型输出值构成的一向量； S305: Using the one-dimensional change training data set obtained in S304, based on the Krging surface fitting algorithm, respectively fitting the case of arbitrarily determined one-dimensional unit variables, the values of the other two unit variables and the corresponding model output values are three-dimensional coordinates, namely ( ,

, ), (

,

,

) (

,

,

) 3 spatial surfaces, each point in the surface is a vector composed of 2-dimensional unit variable values and corresponding model output values;

 个点（），构成

个元素的误差数据训练集合，集合中每个点为一四维向量{

，，

，}，用该误差数据训练集合，训练如图5所示的{

，

，}为3元输入，

为一元输出的神经网络误差修正模型

，并保存于系统用于误差数据修正过程； S306: Among the three spatial surfaces generated in S305, each surface is sampled

points ( ),constitute

The error data training set of elements, each point in the set is a four-dimensional vector{

, ,

, }, use the error data to train the set, and train { as shown in Figure 5

,

, } is a 3-element input,

Neural Network Error Correction Model for Unary Output

, and saved in the system for error data correction process;

As shown in Figure 6, data correction includes the following sub-steps:

S404: Add the final error correction amount to the original data with correction to obtain corrected inclination sensor data.

S401: Regularize each segment of the input data obtained by the data preprocessing module in the data correction step to a length of matrix;

（

 ）（

 ），并将特征值带入训练好的误差修正模型

获得各点数据的误差修正值； S402: Use the same method as S301 to extract the feature value of each point of each piece of data

(

) (

), and the eigenvalues Bring in the trained error correction model

Obtain the error correction value of each point data;

倍上采样形成最终的误差修正量； S403: Connect the error correction values of each point of each segment of data to form a complete error correction value, and perform

Times upsampling to form the final error correction amount;

S404: Add the final error correction amount to the original data with correction to obtain corrected inclination sensor data.

The unit variables in the model training step include frequency h, amplitude f and angle a. the

Claims (3)

1. empirical model measurement of dip angle error information update the system in the wind-tunnel, it is characterized in that: it comprises with lower module:

Data preprocessing module: the input data are carried out pre-service, extract vibrations error character vector, systematic error correction model training process and data correction process all comprise data preprocessing module, but pretreatment module is finished different operating in VEC training process and data correction process;

The model training module: setting up with angle, frequency and amplitude according to the training data that collects in advance and target data is the input variable parameter, error correction values is the joint error correction model of output, and this module comprises that aspect of model vector extracts submodule and model training submodule;

Data correction module: import data according to each section that data preprocessing module in the error correction process obtains, extract the proper vector of each section input data, and proper vector imported the joint error correction model that precondition obtains, obtain error correction values, modified value is added to original input, obtains revised data.

2. the modification method of empirical model measurement of dip angle error information in the wind-tunnel, it is characterized in that: it comprises data pre-service, model training and three steps of data correction, model training and data correction all comprise data preprocessing module, but pretreatment module is finished different operating in model training and data correction;

The data pre-service comprises following substep in the described model training step:

S101: loaded targets data and training data be to update the system, and target data curve and training data curve are carried out down-sampling;

S102: the curve to target data curve and training data carries out wavelet transformation and smoothing processing respectively, obtains their high and low frequency data;

S103: the low-and high-frequency data to the target data curve ask poor with the low-and high-frequency of all training data curves respectively, and input (high frequency poor) and (low frequency poor) output data set of obtaining for training pattern close;

The data pre-service comprises following substep in the described data correction step:

S201: loaded targets data and each wait to revise data to system, and to aim curve and treat that fair curve carries out down-sampling;

S202: to aim curve and treat that fair curve carries out wavelet transformation and smoothing processing, obtain their high and low frequency data respectively;

S203: the low-and high-frequency data to aim curve ask poor with the low-and high-frequency for the treatment of fair curve respectively, obtain input (high frequency poor) data and the reference curve of VEC;

S204: the result according to end-point detection carries out segmentation to the input data that obtain;

Described model training step comprises following substep:

S301: according to the input data that pre-service in the model training step obtains, extract and revise feature;

S302: the output data that obtain with pretreatment module in the model training process are output valve, the correction that obtains among the S301 is characterized as input value, the composing training data acquisition, system ties up as an element variable with each of correction proper vector value, be that the model training of this element variable is gathered with the training data subset division that characterizes each element variable data variation in the training data set, training with each element variable respectively is the input variable parameter, the output data are N (N=element variable number) unit variation model of output valve, and are stored as intermediate result;

S303: point centered by each eigenwert in each unit variable training data subclass that the S302 division obtains, within the specific limits to its interpolation, the new characteristic value collection of each element variable after obtaining respectively to insert;

S304: bring the new characteristic value collection of each element variable that obtains among the S303 into output valve that unit variation model that S302 obtains obtains each model respectively, and with each new element variable characteristic value collection and each model output valve, together constitute one dimension and change the training data set, and storage is used for follow-up model training;

S305: utilize the one dimension that S304 obtains to change the training data set, based on the Krging Algorithm for Surface Fitting, the match M(M=that is three-dimensional coordinate with any two element variable values and corresponding model output valve respectively ) individual space curved surface, the vector of each point for being constituted by 2 dimension element variable values and corresponding model output valve in the curved surface;

S306: K the point of respectively sampling on each curved surface in the M that in S305, the generates space curved surface, constitute the error information training set of KxM element, each point is a four-dimensional vector in the set, with this error information training set training error correction model Q, and the system that is stored in is used for the error information makeover process;

Described data correction comprises following substep:

S401: extract the input data that the data pre-service obtains in the data correction step, and to each input segmentation, organize data into the long certain matrix of row;

S402: adopt same procedure to extract the eigenwert of every segment data with S301, and bring the eigenwert of each segment data into error correction values that the VEC that trains obtains each segment data;

S403: the error correction values of each segment data is connected, form complete error correction values, and carry out up-sampling and form final calibration corrections;

S404: the data addition of final calibration corrections and grandfather tape correction obtains revised obliquity sensor data.

3. the modification method of empirical model measurement of dip angle error information in the wind-tunnel according to claim 2 is characterized in that, the element variable in the described model training step is for comprising frequency h, amplitude f and angle a.

Images