Aerospace equipment data-real fusion test small sample data analysis method

Abstract

The invention discloses an aerospace equipment data-real fusion test small sample data analysis method, which comprises the following steps: designing an aerospace equipment non-destructive data acquisition module which acquires actual operation data through a health management system and is used for extracting health data of aerospace equipment; an aerospace equipment twinborn prediction module is designed, the module takes actual operation data as input, a twinborn learning model is trained to process the data, and a health state prediction result is output; and designing an online transfer learning module which continuously updates and optimizes the twin learning model according to the new operation data. According to the method, on the premise that only a small amount of aerospace equipment information is obtained, the operation health state data of the aerospace equipment can be predicted, the prediction precision is improved through a continuous learning mechanism, and powerful support is provided for maintenance and fault diagnosis of the aerospace equipment.

Description

Technical Field

[0001] The present invention belongs to the fields of electronic engineering and computer science, and specifically relates to a small sample data analysis method for digital-realistic fusion testing of aerospace equipment. Background Art

[0002] Aerospace equipment plays a vital role in the aerospace industry, such as aircraft engines, navigation systems, and flight control systems. If any part of the above systems fails unexpectedly, it will cause a great blow to safety and economy, so it is very important to monitor and diagnose faults in real time and predict the health status with high real-time and accuracy. Since the data collection cost and data volume generated by aerospace equipment during operation are high, how to quickly extract effective information from these limited data and maintain high accuracy and real-time analysis in the case of small sample data is a major challenge faced in the digital-real fusion testing of aerospace equipment. Summary of the invention

[0003] The technical problem to be solved by the present invention is: to provide a method for analyzing small sample data of digital-real fusion testing of aerospace equipment, which covers the design of non-destructive data acquisition module of aerospace equipment, twin prediction module of aerospace equipment and online transfer learning module, and can improve the real-time and accuracy of health status prediction of aerospace equipment to a certain extent, and enhance the safety and economy of aerospace equipment. To this end, the present invention discloses a method for analyzing small sample data of digital-real fusion testing of aerospace equipment, which improves the real-time and accuracy of health status prediction of aerospace equipment to a certain extent, and further ensures the safety and economy of aerospace equipment, through the design of non-destructive data acquisition module of aerospace equipment, twin prediction module of aerospace equipment and online transfer learning module, and with the help of digital twin technology.

[0004] The present invention solves the technical problem by adopting the following technical solution: a method for analyzing small sample data of aerospace equipment digital-real fusion test, comprising:

[0005] Step S1: designing a non-destructive data acquisition module for aerospace equipment, which obtains original equipment data from aerospace equipment and performs data cleaning;

[0006] Step S2: a twin prediction module for aerospace equipment, which first builds a neural network based on physical information, then trains the model with the actual health status data of historical aerospace equipment under normal working conditions, and finally receives the actual health status data of some aerospace equipment under normal working conditions transmitted from step S1 to perform real-time health status prediction;

[0007] Step S3: an online transfer learning module, which is used to realize health status prediction based on small sample data of health status detected in real time under normal working conditions of aerospace equipment.

[0008] The advantages of the present invention compared with the prior art are: with the help of digital twin technology and the use of online physical information networks to obtain historical aerospace equipment health data, network parameters are initialized. On this premise, by obtaining real-time aerospace equipment health data, the health status can be quickly predicted in the case of small sample data. At the same time, the digital twin model iteration is completed based on the loss function output, which can improve the real-time and accuracy of the health status prediction of aerospace equipment to a certain extent, and improve the safety and economy of aerospace equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a flowchart of a small sample data analysis method for digital-realistic fusion testing of aerospace equipment according to the present invention. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in each embodiment of the present invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above-mentioned purpose, the present invention adopts the following technical scheme.

[0011] The present invention relates to a small sample data analysis method for digital-real fusion testing of aerospace equipment, including the design of an aerospace equipment non-destructive data acquisition module, an aerospace equipment twin prediction module and an online transfer learning module, which can improve the real-time and accuracy of the health status prediction of aerospace equipment to a certain extent, and enhance the safety and economy of aerospace equipment.

[0012] The flowchart of the present invention is as follows Figure 1 As shown, taking the engine in aerospace equipment as an example, the specific implementation is as follows:

[0013] Step S1, designing an engine non-destructive data acquisition module, which obtains raw device data from the engine and performs data cleaning, and is specifically implemented as follows:

[0014] In view of the difference in health data status between the engine under normal operating conditions and when it is not in use, a health management system is designed. The system ensures that real-time health data can be continuously collected at all working stages of the engine through sensor networks, data acquisition cards and communication interface facilities.

[0015] In view of the complex engine working conditions and large noise interference, a data cleaning module is designed. The outlier detection algorithm is used to first calculate the quartiles. and interquartile range : =25th percentile, =75th percentile, , then determine the outlier limit: Lower bound = , upper bound = , and finally determine the outliers for all data: if Nether or The upper bound is Outliers are eliminated to achieve deep purification of the collected data and ensure that the health status data used in subsequent analysis is reliable.

[0016] In case of small and complex engine operating condition data, the Pearson correlation coefficient of the data is calculated: ,in and There are two characteristics, is the sample point In Features The value below, is the sample point In Features The value below, and The samples are and The mean value under It is a feature and Features The Pearson correlation coefficient between is the number of samples. The correlation between the features and health status is compared by comparing the size of the coefficient values ​​and the physical meaning of the parameters is considered for feature selection. For each feature's peak value, mean, standard deviation, kurtosis, curve entropy, and curve slope, the principal component analysis method is used: First, the covariance matrix of the data is calculated. , and then solve the covariance matrix The eigenvalues ​​and corresponding eigenvectors of the data are obtained, where the eigenvalues ​​reflect the variances in each feature direction and the eigenvectors indicate the directions of these variances. Finally, the eigenvalues ​​are arranged in descending order and the eigenvectors corresponding to the first three largest eigenvalues ​​are selected as the principal components to further extract and strengthen the features that are highly correlated with the health status, so as to improve the accuracy and efficiency of data analysis.

[0017] The calculation formula for kurtosis is: ,in It is the first The observed value of sample points, is the mean of the entire sample, is the number of sample points, is the variance of the entire sample;

[0018] The calculation formula of curve entropy is: ,in Indicates the data set The normalized value of the data point, Represents the total number of data points in the dataset;

[0019] The calculation formula for the slope of the curve is: ,in represents the observed value at the start time of a cycle, represents the observed value at the end of a cycle, Indicates the total time from the start time to the end time of a cycle.

[0020] Step S2, designing an engine twin prediction module, which first constructs a neural network based on physical information, where the physical information includes structural mechanics properties, material properties, and fluid dynamics properties, and then uses historical engine normal working condition measured health status data for model training, and finally receives part of the engine normal working condition measured health status data transmitted in step S1 for real-time health status prediction, which is specifically implemented as follows:

[0021] Combined with the actual physical characteristics of the engine, specifically, a neural network model including a health state decay rate model and a health state decay model is built. The input of the network is the feature data screened out in step S1. The health state decay rate is calculated through two layers of MLP and one layer of FC. Then the predicted health state decay rate and the data screened out by S1 are passed into the second part of the model composed of one layer of MLP and one layer of FC to predict the health status of the aerospace equipment.

[0022] MLP is a multi-layer perceptron, a feedforward neural network that contains three layers of nodes: an input layer, a hidden layer, and an output layer. Each node (except the input node) is a neuron with a nonlinear activation function. Through the full connection between these layers, the input samples are fed forward layer by layer to finally output the high-level abstract information extracted from the data;

[0023] FC is a fully connected layer, in which each neuron is connected to all neurons in the previous layer. Each connection corresponds to a learnable weight, which can integrate the information from the MLP layer and output the final prediction result.

[0024] After the above network is built, the model parameters are initialized, and then the neural network model is trained using the measured health status data under normal engine conditions. The specific training method is as follows: First, the input data is feedforward calculated, and the Euclidean distance between the predicted health status decay rate and the true value is calculated after each round of feedforward operation: ,in is the predicted output of the network is the true value, and the calculated result is used as the first penalty term; the Euclidean distance between the predicted health state and the true health state is used as the second penalty term, and the sum of the two penalty terms is used as the total penalty term, and then the back propagation algorithm is used: ,in is the learning rate, are model parameters, is the total penalty term, which adjusts the model parameters to ensure that the model can accurately simulate the changes in the health status of the engine;

[0025] Step S3, designing an online transfer learning module, which is used to realize the health status prediction based on the real-time detection of the health status of the engine under normal working conditions. The specific implementation is as follows:

[0026] Input the engine real-time feature small sample data after outlier processing, feature selection and principal component analysis in step S1 into the neural network framework whose parameters have been updated by historical data in step S2, so as to output the engine health status predicted by the network;

[0027] The predicted real-time health status of the engine is compared with the actual health status data obtained in step S1 to calculate the loss function, and the parameters of the physical information network are fine-tuned through the back propagation algorithm according to the result of the loss function to improve the prediction performance and generalization ability of the model on small sample data sets.

[0028] In summary, the present invention discloses a small sample data analysis method for digital-real fusion testing of aerospace equipment, including the design of an aerospace equipment non-destructive data acquisition module, an aerospace equipment twin prediction module and an online transfer learning module, which can improve the real-time and accuracy of the health status prediction of aerospace equipment to a certain extent, and enhance the safety and economy of aerospace equipment.

[0029] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in this field.

[0030] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A small sample data analysis method for digital-real fusion testing of aerospace equipment, characterized in that: include: Step S1: designing a non-destructive data acquisition module for aerospace equipment, which obtains original equipment data from aerospace equipment and performs data cleaning; Step S2: Design an aerospace equipment twin prediction module, which first builds a neural network based on physical information, then uses historical aerospace equipment measured health status data under normal working conditions to train the model, and finally receives some of the aerospace equipment measured health status data under normal working conditions transmitted in step S1 to perform real-time health status prediction; Step S3: Design an online transfer learning module, which is used to realize health status prediction based on small sample data of health status detected in real time under normal working conditions of aerospace equipment.

2. According to claim 1, a method for analyzing small sample data of aerospace equipment digital-real fusion test is characterized in that: The non-destructive data acquisition module of aerospace equipment in step S1 is specifically implemented as follows: Step S1.1: Design a health management system to address the differences in health data status between aerospace equipment in normal working conditions and when not in use. The system uses sensor networks, data acquisition cards, and communication interface facilities to ensure that real-time health data is collected continuously at all working stages of aerospace equipment. Step S1.2: Design a data cleaning module to address the complex working conditions of aerospace equipment and the large noise interference. Use the outlier detection algorithm to first calculate the quartiles. and interquartile range : =25th percentile, =75th percentile, , then determine the outlier limit: Lower bound = , upper bound = , and finally determine the outliers for all data: if Nether or The upper bound is Outliers are removed to achieve deep purification of the collected data; Step S1.3: For the case where the amount of aerospace equipment working condition data is small and complex, calculate the Pearson correlation coefficient of the data: ,in and There are two characteristics, is the sample point In Features The value below, is the sample point In Features The value below, and The samples are characterized by and The mean value under It is a feature and Features The Pearson correlation coefficient between is the number of samples. The correlation between the features and health status is compared by comparing the values ​​of the Pearson correlation coefficient and the physical meaning of the parameters is considered for feature selection. The peak value, mean, standard deviation, kurtosis, curve entropy and curve slope of each feature are analyzed by principal component analysis, that is, the covariance matrix of the data is first calculated. , and then solve the covariance matrix The eigenvalues ​​and corresponding eigenvectors of are obtained, where the eigenvalues ​​reflect the variances in each feature direction and the eigenvectors indicate the directions of these variances. Finally, the eigenvalues ​​are arranged in descending order and the top The eigenvector corresponding to the largest eigenvalue is used as the principal component to further extract and strengthen the features that are highly related to the health status to improve the accuracy and efficiency of data analysis; The formula for kurtosis is: ,in It is the first The observed value of sample points, is the mean of the entire sample, is the number of sample points, is the variance of the entire sample; The formula for curve entropy is: ,in Indicates the data set The normalized value of the data point, Represents the total number of data points in the dataset; The formula for the slope of the curve is: ,in represents the observed value at the start time of a cycle, represents the observed value at the end of a cycle, Indicates the total time from the start time to the end time of a cycle.

3. The method for analyzing small sample data of aerospace equipment digital-real fusion test according to claim 2 is characterized in that: The specific implementation of the aerospace equipment twin prediction module in step S2 is as follows: Step S2.1: In combination with the actual physical characteristics of aerospace equipment, specifically, a neural network model including a health state decay rate model and a health state decay model is built. The input of the network is the feature data screened in step S1. The health state decay rate is calculated through two layers of MLP and one layer of FC. Then, the predicted health state decay rate and the data screened by S1 are passed together to the second part of the model consisting of one layer of MLP and one layer of FC to predict the health state of the aerospace equipment. MLP is a multi-layer perceptron, a feedforward neural network that consists of three layers of nodes: an input layer, a hidden layer, and an output layer. Each node, except the input node, is a neuron with a nonlinear activation function. Through the full connection between these layers, the input samples are fed forward layer by layer to finally output the high-level abstract information extracted from the data. FC is a fully connected layer, in which each neuron is connected to all neurons in the previous layer. Each connection corresponds to a learnable weight, which can integrate the information from the MLP layer and output the final prediction result. After building the above network, the model parameters are initialized. Step S2.2: Use the measured health status data of aerospace equipment under normal working conditions to train the neural network model. The specific training method is as follows: First, the input data is fed forward, and the Euclidean distance between the predicted health status decay rate and the true value is calculated after each round of feedforward operation: ,in is the predicted output of the network is the true value, and the calculated result is used as the first penalty term; the Euclidean distance between the predicted health state and the true health state is used as the second penalty term, and the sum of the two penalty terms is used as the total penalty term, and then the back propagation algorithm is used: ,in is the learning rate, are model parameters, is the total penalty term, which is used to adjust the model parameters to ensure that the model can accurately simulate the changes in the health status of aerospace equipment.

4. The method for analyzing small sample data of aerospace equipment digital-real fusion test according to claim 3 is characterized in that: The online transfer learning module of step S3 is specifically implemented as follows: Step S3.1: input the real-time characteristic small sample data of aerospace equipment obtained after processing in step S1 into the neural network model whose parameters have been updated by historical data in step S2, so as to output the real-time health status of aerospace equipment predicted by the network; Step S3.2: Compare the predicted real-time health status of aerospace equipment with the actual health status data obtained to calculate the loss function, and fine-tune the parameters of the physical information network through the back propagation algorithm based on the result of the loss function to improve the prediction performance and generalization ability of the model on small sample data sets.

Images