Avionics system fault prediction method based on transfer learning abnormal working condition

Abstract

The invention relates to the field of avionics system fault prediction, in particular to an avionics system fault prediction method based on a transfer learning abnormal working condition, and the method specifically comprises the following steps: S1, obtaining multi-source heterogeneous data in an avionics system and a source system in real time, and carrying out the preprocessing of the multi-source heterogeneous data, and generating first data; s2, performing feature extraction on the first data to obtain key features related to the fault evolution process; s3, taking the historical key features corresponding to the source system as training samples, making sample tags according to the corresponding fault tags, and constructing a training set; through the transfer learning technology, fault data of an existing system can be transferred to a target system, knowledge sharing and reutilization are achieved, the prediction capacity and adaptability of the model in a new environment are greatly improved, and particularly under the condition that the fault data are limited, high-accuracy fault prediction can still be provided.

Description

Technical Field

[0001] The present invention relates to the field of avionics system fault prediction, and specifically to a fault prediction method for avionics systems under abnormal conditions based on transfer learning. Background Art

[0002] With the increasing complexity of avionics systems, fault diagnosis and prediction have become key technologies to ensure the safe and reliable operation of aircraft. Traditional fault diagnosis methods mainly rely on rule-based expert systems, model prediction, and experience-based fault databases. These methods usually rely on a large amount of historical data and prior knowledge. However, when dealing with complex, multi-fault, and dynamically changing avionics systems, these methods are prone to problems such as low diagnostic accuracy and slow response speed, and the accuracy of predicting faults in avionics systems is relatively low in the absence of a large amount of historical data. Summary of the Invention

[0003] Aiming at the deficiencies of the prior art, the present invention provides a fault prediction method for avionics systems under abnormal conditions based on transfer learning, which solves the technical problem of relatively low accuracy in predicting faults in avionics systems due to the lack of a large amount of historical data in the prior art.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A fault prediction method for avionics systems under abnormal conditions based on transfer learning, specifically including the following steps:

[0006] S1. Real-time obtain multi-source heterogeneous data in the avionics system and the source system, and preprocess the multi-source heterogeneous data to generate first data;

[0007] S2. Extract features from the first data to obtain key features related to the fault evolution process;

[0008] S3. Use the historical key features corresponding to the source system as training samples, make sample labels according to the corresponding fault labels, and construct a training set;

[0009] S4. Train a preliminary model according to the training set to obtain a preliminary fault prediction model;

[0010] S5. Construct a target prediction model according to the preliminary fault prediction model;

[0011] S6. Input the current key features of the avionics system into the target model, and output the probability of each type of fault occurring in the avionics system.

[0012] Further, in step S1, it specifically includes the following steps:

[0013] S11. Normalize the multi-source heterogeneous data to generate the second data, and its expression is:

[0014]

[0015] In the formula, X 1 represents the data point after normalization; X 1 represents the original data point; μ1 represents the average value of each type of data in the multi-source heterogeneous data; σ 1 represents the standard deviation of each type of data in the multi-source heterogeneous data;

[0016] S12. Fill in the missing items in the second data to generate the first data, and its expression is:

[0017]

[0018] In the formula, y represents the estimated missing value; x represents the position of the missing value; x 1 , y 1 and x 2 , y 2 respectively represent the known data on both sides of the missing value.

[0019] Furthermore, in step S2, it specifically includes the following steps:

[0020] S21. Extract the basic statistical features of the first data. The basic statistical features include the mean feature μ 2 , variance feature σ 2 2 , skewness feature Skewness and kurtosis feature Kurtosis, and their calculation formulas are respectively:

[0021]

[0022] In the formula, N represents the total number of samples of the first data; x i represents the i-th sample point;

[0023]

[0024]

[0025]

[0026] S22. Extract the signal morphology features of the first data. The signal morphology features include rise time, fall time, number of wave peaks, and number of wave valleys;

[0027] S23. Extract the frequency domain features of the first data. The frequency domain features include fast Fourier transform and power spectral density;

[0028] S24. Extract the time-domain features of the first data;

[0029] S25. Make fault labels according to each occurred fault type and level, and calculate the correlation coefficient between each feature and the fault labels

[0030] S26. Select the correlation coefficient The highest several features as the key features.

[0031] Further, in step S23, the expression of the fast Fourier transform is:

[0032]

[0033] In the formula, X k represents the k-th component of the n-th sample in the frequency domain in the first data; x n represents the n-th sample in the time domain; j represents the imaginary unit; k is the frequency index, representing different frequency components;

[0034] The expression of the power spectral density is:

[0035] P(f) = |X(f)| 2 / N

[0036] In the formula, P(f) represents the power spectral density at the frequency f; |X(f)| represents the frequency-domain amplitude at the frequency f.

[0037] Further, in step S24, the time-domain features of the first data are extracted by wavelet transform, and its expression is:

[0038]

[0039] In the formula, W a (j, k) represents the approximation coefficient of the j-th layer; W d (j, k) represents the detail coefficient of the j-th layer; x[n] represents the input of the first data; k is the position index, representing the coefficients at different positions; j represents the decomposition level; φ(n) and ψ(n) respectively represent the scaling function and the wavelet basis function;

[0040] The iterative formulas of φ(n) and ψ(n) are:

[0041]

[0042] In the formula, h[m] represents the low-pass filter coefficient; g[m] represents the high-pass filter coefficient; m represents the discrete-time index.

[0043] Further, in step S25, it specifically includes the following steps:

[0044] S251. Optionally select one feature as the target feature X i , and the remaining features are (Z 1 , …, Z m );

[0045] S252. Construct a multiple linear regression model, and its expression is:

[0046] X i = β 0 + β 1 Z 1 + … + β m Z m + ∈

[0047] In the formula, β 0 represents the intercept term; β 1 , …, β m respectively represent the regression coefficients of Z 1 , …, Z m ; ∈ represents the error term;

[0048] S253. Calculate the predicted value i of X according to the multiple linear regression model Calculate the residual eX i between X and i ;

[0049] S254. Repeat steps S252 - S253 to calculate the residual eY of the fault label;

[0050] S255. Calculate the correlation coefficient Its calculation formula is:

[0051]

[0052] In the formula, eX i，j represents the residual of X j corresponding to the j-th feature; eY j represents the residual of the fault label corresponding to X j ; and respectively represent the average values of all residuals eX i and eY.

[0053] Furthermore, in step S4, the preliminary model includes a first convolutional layer, a second convolutional layer, an LSTM layer, and a fully connected layer;

[0054] The first convolutional layer inputs the key features and outputs the first updated features;

[0055] The second convolutional layer inputs the first updated features and outputs the second updated features;

[0056] The LSTM layer takes the second updated feature as input and outputs the third updated feature;

[0057] The fully connected layer takes the third updated feature as input and outputs the occurrence probability of each fault type.

[0058] Further, in step S5, it specifically includes the following steps:

[0059] S51. Use the historical key features of the avionics system as training samples, make sample labels according to the corresponding fault labels, and construct a data set based on the training samples and sample labels;

[0060] S52. Construct an original target prediction model according to the preliminary fault prediction model. The architecture of the original target prediction model is the same as that of the preliminary fault prediction model, and the parameters in the first convolutional layer, the second convolutional layer, and the LSTM layer are consistent;

[0061] S53. Use the data set to train the original target prediction model to obtain the target prediction model.

[0062] Compared with the prior art, the present invention provides a method for predicting faults in an avionics system under abnormal conditions based on transfer learning, and has the following beneficial effects:

[0063] 1. In the present invention, through the transfer learning technology, the invention can transfer the fault data of the existing system to the target system, realize the sharing and reuse of knowledge, greatly improve the prediction ability and adaptability of the model in the new environment, and especially can still provide high-accuracy fault prediction in the case of limited fault data.

[0064] 2. The present invention breaks through the limitations of the existing methods in dealing with diverse and complex data sources through the multi-source heterogeneous data fusion technology. The data collected by different sensors and devices often have heterogeneity, while the present invention can integrate data from multiple sources through an innovative data fusion mechanism, effectively extract key fault features, and enhance the reliability and accuracy of fault prediction.

[0065] 3. In the present invention, the first data is feature-extracted in various ways to ensure that the extracted features are comprehensive enough to facilitate more accurate fault prediction of the avionics system in the later stage. And when calculating the correlation coefficient, the indirect influence of other features on the correlation coefficient is considered, effectively improving the accuracy of calculating the correlation coefficient, and further improving the accuracy of selecting key features. Description of the Drawings

[0066] The drawings described herein are used to provide a further understanding of the present application, and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation to the present application. In the drawings:

[0067] Figure 1 This is a flowchart of a fault prediction method for an avionics system under abnormal conditions based on transfer learning of the present invention. Detailed implementation manners

[0068] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners. Thereby, the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0069] With the increasing complexity of avionics systems, their fault diagnosis and prediction have become key technologies to ensure the safe and reliable operation of aircraft. An avionics system usually consists of multiple modular subsystems, and the information interaction between subsystems is complex and highly dependent. Faults may not only affect the normal operation of a single component, but also spread through system coupling, resulting in multi-point faults, thereby endangering the safety of the aircraft. Therefore, how to accurately and timely identify and predict faults has become an important technical challenge for avionics systems.

[0070] Currently, the fault diagnosis and prediction technology for avionics systems has made some progress to a certain extent, but there are still significant limitations. Traditional fault diagnosis methods mainly rely on rule-based expert systems, model prediction, and experience-based fault databases. These methods usually rely on a large amount of historical data and prior knowledge. However, when dealing with complex, multi-fault, and dynamically changing avionics systems, these methods are prone to problems such as low diagnostic accuracy and slow response speed. In terms of the processing of multi-source heterogeneous data, existing technologies are difficult to effectively integrate data from different sensors, devices, and monitoring systems, resulting in information loss or insufficient analysis, limiting the accuracy of fault prediction.

[0071] In addition, existing fault prediction methods often rely on static models or overly simplified assumptions, and it is difficult to effectively predict faults in real time under complex working conditions. Especially in the rapidly changing working conditions of avionics systems, traditional models do not fully consider the complexity of system dynamic evolution, resulting in poor timeliness and accuracy of fault prediction. At the same time, due to the scarcity and diversity of avionics system fault data, many traditional algorithms cannot handle situations of insufficient data or cross-domain, resulting in insufficient generalization ability of fault prediction.

[0072] In view of the above deficiencies in the prior art, the present invention proposes a fault prediction method for an avionics system based on transfer learning, as Figure 1 shown, which specifically includes the following steps:

[0073] S1. Obtain multi-source heterogeneous data in the avionics system and the source system in real time, and preprocess the multi-source heterogeneous data to generate the first data. Specifically, the multi-source heterogeneous data includes:

[0074] Temperature information: Use temperature sensors to monitor the temperature changes of key components or the environment, which helps to identify overheating or other abnormal conditions;

[0075] Pressure information: Use pressure sensors to measure the pressure levels of hydraulic systems and pneumatic systems, and help detect problems such as leaks or blockages;

[0076] Vibration information: Use vibration sensors to capture the vibration patterns of mechanical equipment, which is very important for early detection of mechanical failures (such as bearing wear);

[0077] Acceleration information: Use accelerometers to record the acceleration information of objects, which can be used to analyze the attitude and motion state of the aircraft;

[0078] Rotation information: Use gyroscopes to obtain angular velocity information for attitude control and navigation;

[0079] Circuit information: Use voltage and current sensors to monitor the voltage and current in the circuit to ensure the stability and safety of the power system;

[0080] Power factor: Reflects the efficiency of electrical energy use, and abnormal values may indicate potential problems;

[0081] ARINC 429 / 664 (AFDX): Transmits real-time data between avionics systems, including information such as flight control systems and engine management systems;

[0082] Flight recorder (black box): Stores various operation parameters and system states during flight, and is an important basis for accident investigation;

[0083] Maintenance log: Records previous maintenance activities and provides support for historical data analysis;

[0084] Camera: Installed outside or inside the aircraft to monitor the status of specific areas, such as the retraction and extension of the landing gear;

[0085] Infrared camera: Can detect the heat distribution invisible to the naked eye and is suitable for detecting hidden fault points;

[0086] Meteorological data: External weather conditions have a direct impact on flight safety, so factors such as wind speed and humidity need to be considered;

[0087] Geographic Information System (GIS) data: Provides context information related to geographical locations, such as the topographic and geomorphic features required for flight path planning;

[0088] Flight simulator output: By simulating the system responses under different flight scenarios, a dataset for training and validating the model is generated.

[0089] The multi-source heterogeneous data constructed based on the above information can reflect the operating conditions of the avionics system and the source system in many aspects, facilitating the accuracy guarantee when evaluating and predicting the faults of the avionics system in the later stage.

[0090] In addition, after obtaining the above multi-source heterogeneous data, since the numerical ranges of each data are not unified and there are missing items, which will affect the subsequent prediction results of faults. Therefore, in step S1, it specifically includes the following steps:

[0091] S11. Normalize the multi-source heterogeneous data to generate the second data, and its expression is:

[0092]

[0093] In the formula, X 1 represents the normalized data point; X 1 represents the original data point; μ1 represents the average value of each data in the multi-source heterogeneous data; σ 1 represents the standard deviation of each data in the multi-source heterogeneous data;

[0094] S12. Fill in the missing items of the second data to generate the first data, and its expression is:

[0095]

[0096] In the formula, y represents the estimated missing value; x represents the position of the missing value; x 1 , y 1 and x 2 , y 2 respectively represent the known data on both sides of the missing value. After processing the multi-source heterogeneous data, the problem of inconsistent data and missing items can be avoided from affecting the prediction accuracy of faults.

[0097] S2. Extract the key features related to the fault evolution process from the first data; specifically, the faults that occur during the operation of the avionics system and the source system will be reflected in the first data. Different fault types and different fault levels will cause changes in different types of the first data. From the above content, it can be seen that the fault information of the avionics system can be calculated based on the key features of the first data. Therefore, in step S2, it specifically includes the following steps:

[0098] S21. Extract the basic statistical features of the first data, and the basic statistical features include the mean feature μ 2 , the variance feature σ 22 The skewness feature Skewness and kurtosis feature Kurtosis, and their calculation formulas are respectively:

[0099]

[0100] In the formula, N represents the total number of samples of the first data; x i represents the i-th sample point;

[0101]

[0102]

[0103]

[0104] The basic statistical features reflect the overall distribution characteristics of the signal. Extracting these features can describe the average level, fluctuation degree, symmetry, and kurtosis of the data distribution of the signal.

[0105] S22. Extract the signal shape features of the first data. The signal shape features include rise time, fall time, number of wave peaks, and number of wave valleys. Specifically, the rise time represents the time required for the signal to rise from a certain low level to a high level; the fall time represents the time required for the signal to fall from a high level to a low level; the number of wave peaks represents the number of maximum values that the signal appears within a certain period of time; the number of wave valleys represents the number of minimum values that the signal appears within a certain period of time, which helps to capture the change trend and abnormal fluctuations of the signal, and further identify potential fault patterns;

[0106] S23. Extract the frequency domain features of the first data. The frequency domain features include fast Fourier transform and power spectral density. Specifically, in step S23, the expression of the fast Fourier transform is:

[0107]

[0108] In the formula, X k represents the k-th component of the n-th sample in the frequency domain of the first data; x n represents the n-th sample in the time domain; j represents the imaginary unit; k is the frequency index, representing different frequency components. Specifically, the time domain signal is converted into a frequency domain signal through the fast Fourier transform to analyze the energy distribution of different frequency components; spectral analysis can reveal the periodic components and noise level existing in the signal, and help to identify specific frequency features related to faults;

[0109] The expression of the power spectral density is:

[0110] P(f) = |X(f)| 2 / N

[0111] Wherein, P(f) represents the power spectral density at frequency f; |X(f)| represents the frequency domain amplitude at frequency f. Specifically, the power distribution of the signal at each frequency is calculated to evaluate the energy concentration of the signal. Power spectral density analysis can quantify the energy contributions of different frequency components, thereby identifying the abnormal energy distribution caused by faults.

[0112] S24. Extract the time-domain features of the first data; specifically, in step S24, the time-domain features of the first data are extracted through wavelet transform, and its expression is:

[0113]

[0114] Wherein, W a (j, k) represents the approximation coefficient of the j-th layer; W d (j, k) represents the detail coefficient of the j-th layer; x[n] represents the input of the first data; k is the position index, representing the coefficients at different positions; j represents the decomposition level; φ(n) and ψ(n) respectively represent the scaling function and the wavelet basis function;

[0115] The iterative formulas for φ(n) and ψ(n) are:

[0116]

[0117] Wherein, h[m] represents the low-pass filter coefficient; g[m] represents the high-pass filter coefficient; m represents the discrete-time index.

[0118] S25. Make a fault label according to each occurred fault type and level, and calculate the correlation coefficient between each feature and the fault label Specifically, the common feature selection method is to calculate the correlation coefficient between each feature and the fault after feature selection, and then select several features with the highest values as the key features. However, this method does not consider the influence of the final correlation coefficient caused by the interaction between different features, resulting in inaccurate calculation results. Therefore, in step S25, it specifically includes the following steps:

[0119] S251. Arbitrarily select a feature as the target feature X i , and the remaining features are (Z 1 , …, Z m );

[0120] S252. Construct a multiple linear regression model, and its expression is:

[0121] X i = β 0 + β 1 Z 1 + … + β m Z m + ∈

[0122] In the formula, β 0 represents the intercept term; β 1 , …, β m respectively represent the regression coefficients of Z 1 , …, Z m ; ∈ represents the error term;

[0123] S253. Calculate the predicted value of X i according to the multiple linear regression model Calculate the residual eX between X i and ; i ;

[0124] S254. Repeat steps S252 - S253 to calculate the residual eY of the fault label;

[0125] S255. Calculate the correlation coefficient The calculation formula thereof is:

[0126]

[0127] In the formula, eX i，j represents the residual of X corresponding to the j-th feature j ; eY j represents the residual of the fault label corresponding to X j ; and respectively represent the average values of all residuals eX i and eY; Specifically, review the past maintenance logs, mark the time periods during which faults occurred, and the specific types of faults. If the equipment operates normally during a certain time period, the label for that time period is "no fault"; if a fault occurs, mark the corresponding fault code as the fault label.

[0128] S26. Select several features with the highest correlation coefficient as key features.

[0129] In step S2 of the present invention, when extracting the features of the first data, the time domain features, frequency domain features, and time-frequency domain features are extracted, the change trend and abnormal fluctuations of the signal are captured, and the periodic components and energy distribution existing in the signal are explained. Moreover, the method of wavelet transform is used to extract the information of time and frequency simultaneously, which is convenient for identifying transient faults and persistent faults. In step S2 of the present invention, the first data is subjected to feature extraction in various ways to ensure that the extracted features are comprehensive enough, which is convenient for more accurate fault prediction of the avionics system in the later stage. And when calculating the correlation coefficient, the indirect influence of other features on the correlation coefficient is taken into account, effectively improving the accuracy of calculating the correlation coefficient, and further improving the accuracy of selecting key features.

[0130] Through the multi-source heterogeneous data fusion technology, the present invention breaks through the limitations of existing methods in dealing with diverse and complex data sources. The data collected by different sensors and devices often has heterogeneity, while the present invention can integrate data from multiple sources through an innovative data fusion mechanism, effectively extract key fault features, and enhance the reliability and accuracy of fault prediction.

[0131] S3. Use the historical key features corresponding to the source system as training samples, make sample labels according to the corresponding fault labels, and construct a training set. Specifically, in order to improve the robustness of the preliminary fault prediction model and avoid overfitting or underfitting, it is necessary to train the preliminary model. In the avionics system, due to the low occurrence frequency of faults and the different specific situations of each fault, the amount of fault data for training the model is very limited. Traditional machine learning methods usually require a large amount of labeled data to train a well-performing model, but it is difficult to obtain enough data in this case. Therefore, in the present invention, the fault types and multi-source heterogeneous data accumulated in other systems (source systems) can be obtained, and a training set can be constructed based on them to improve the robustness of the preliminary fault prediction model.

[0132] It should be noted that a one-hot encoding is assigned to each fault type according to the fault label, and each one-hot encoding corresponds to a sample label.

[0133] S4. Train the preliminary model according to the training set to obtain a preliminary fault prediction model. As a further implementation manner of the present invention, after obtaining the key features, in order to enable the preliminary fault prediction model to accurately output fault labels according to the key features, it is necessary for the preliminary fault prediction model to fully extract the information in the key features. Therefore, in step S4, the preliminary model includes a first convolutional layer, a second convolutional layer, an LSTM layer, and a fully connected layer;

[0134] The first convolutional layer inputs the key features and outputs the first updated features;

[0135] The second convolutional layer inputs the first updated features and outputs the second updated features;

[0136] The LSTM layer inputs the second updated features and outputs the third updated features;

[0137] The fully connected layer inputs the third updated features and outputs the occurrence probability of each fault type. Specifically, through appropriate output layer configuration of the preliminary model, the model can provide accurate fault prediction results. This architecture not only considers the temporal correlation of features but also can adapt to complex dynamic working conditions, thus improving the accuracy and real-time performance of fault prediction.

[0138] S5. Construct a target prediction model based on the preliminary fault prediction model. Specifically, the preliminary fault prediction model is trained according to the key features of the source system. Due to the differences in factors such as the functions, structural designs, and usage conditions of different systems, the preliminary fault prediction model cannot directly predict the probability of faults in the avionics system based on the key features of the avionics system, resulting in errors. Therefore, in step S5, the following steps are specifically included:

[0139] S51. Use the historical key features of the avionics system as training samples, make sample labels according to the corresponding fault labels, and construct a data set based on the training samples and sample labels.

[0140] S52. Construct an original target prediction model based on the preliminary fault prediction model. The architecture of the original target prediction model is the same as that of the preliminary fault prediction model, and the parameters in the first convolutional layer, the second convolutional layer, and the LSTM layer are consistent.

[0141] S53. Train the original target prediction model using the data set to obtain the target prediction model. In step S5 of the present invention, when performing transfer learning, the architecture of the preliminary fault prediction model and the parameters in the first convolutional layer, the second convolutional layer, and the LSTM layer are retained, and only the historical key features of the avionics system are used to train the original target prediction model to update the parameters in the fully connected layer, so that the target prediction model can adapt to specific data distributions and operating environments with less training data.

[0142] In the present invention, through the transfer learning technology, the invention can transfer the fault data of the existing system to the target system, realize the sharing and reuse of knowledge, greatly improve the prediction ability and adaptability of the model in the new environment, and still provide high-accuracy fault prediction especially when the fault data is limited.

[0143] S6. Input the current key features of the avionics system into the target model, and output the probability of each type of fault occurring in the avionics system. Specifically, when the target model outputs the fault probability of the avionics system, the staff can take corresponding measures, such as maintenance and repair operations, in the shortest time according to the fault probability, which can effectively reduce losses.

[0144] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above embodiment methods can be completed by instructing relevant hardware through a program. Therefore, this application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0145] The above embodiments have introduced the present invention in detail. Specific examples are used in this article to elaborate on the principle and implementation of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for predicting avionics system faults under abnormal conditions based on transfer learning, characterized in that: The specific steps include: S1. Acquire multi-source heterogeneous data in the avionics system and the source system in real time, and pre-process the multi-source heterogeneous data to generate first data; S2. Extracting features from the first data to obtain key features related to the fault evolution process; S3. Use the historical key features corresponding to the source system as training samples, create sample labels according to the corresponding fault labels, and construct a training set; S4. Training the preliminary model according to the training set to obtain a preliminary fault prediction model; S5. construct a target prediction model based on the preliminary fault prediction model; S6. Input the current key characteristics of the avionics system into the target model and output the probability of each type of failure occurring in the avionics system.

2. The avionics system fault prediction method according to claim 1, characterized in that: In step S1, the following steps are specifically included: S11, normalize the multi-source heterogeneous data to generate second data, the expression of which is: Where, X 1 represents the normalized data point; X 1 represents the original data point; μ1 represents the average value of each data in the multi-source heterogeneous data; σ1 represents the standard deviation of each data in the multi-source heterogeneous data; S12, fill the missing items of the second data to generate the first data, which is expressed as: In the formula, y represents the estimated missing value; x represents the location of the missing value; x1, y1 and x2, y2 represent the known data on both sides of the missing value respectively.

3. The avionics system fault prediction method according to claim 1, characterized in that: In step S2, the following steps are specifically included: S21, extracting basic statistical features of the first data, the basic statistical features including mean feature μ2, variance feature σ2 2 , Skewness and Kurtosis, the calculation formulas are: Where N represents the total number of samples of the first data; x i represents the i-th sample point; S22, extracting signal morphology features of the first data, where the signal morphology features include a rise time, a fall time, a number of peaks, and a number of troughs; S23, extracting frequency domain features of the first data, where the frequency domain features include fast Fourier transform and power spectral density; S24, extracting time domain features of the first data; S25. Create a fault label based on the type and level of each fault that has occurred, and calculate the correlation coefficient between each feature and the fault label S26. Select correlation coefficient The highest number of features are used as key features.

4. The avionics system fault prediction method according to claim 3, characterized in that: In step S23, the expression of fast Fourier transform is: Where, X k represents the kth component of the nth sample in the first data in the frequency domain; x n represents the nth sample in the time domain; j represents the imaginary unit; k is the frequency index, representing different frequency components; The expression of power spectral density is: P(f)=|X(f)| 2 / N Where P(f) represents the power spectral density at frequency f; |X(f)| represents the frequency domain amplitude at frequency f.

5. The avionics system fault prediction method according to claim 3, characterized in that: In step S24, the time domain features of the first data are extracted by wavelet transform, and the expression is: Where W a (j, k) represents the approximation coefficient of the jth layer; W d (j, k) represents the detail coefficient of the jth layer; x[n] represents the input of the first data; k is the position index, representing the coefficients at different positions; j represents the decomposition level; φ(n) and ψ(n) represent the scaling function and the wavelet basis function respectively; The iterative formulas for φ(n) and ψ(n) are: Wherein, h[m] represents the low-pass filter coefficient; g[m] represents the high-pass filter coefficient; and m represents the discrete time index.

6. The avionics system fault prediction method according to claim 3, characterized in that: In step S25, the following steps are specifically included: S251. Select one feature as the target feature X i , the remaining features are (Z1,…,Z m ); S252. Construct a multiple linear regression model, whose expression is: X i =β0+β1Z1+…+β m Z m +∈ Where β0 represents the intercept term; β1, …, β m Respectively represent Z1, ..., Z m The regression coefficient of ;∈ represents the error term; S253, calculate X according to the multivariate linear regression model i The predicted value of Calculate X i and The residual eX i ; S254, repeat steps S252-S253 to calculate the residual eY of the fault label; S255. Calculate the correlation coefficient The calculation formula is: In the formula, eX i，j Indicates X corresponding to the jth feature j The residual of eY j Represents X j The residual of the corresponding fault label; and Represent all residuals eX respectively i and the average value of eY.

7. The avionics system fault prediction method according to claim 1, characterized in that: In step S4, the preliminary model includes a first convolutional layer, a second convolutional layer, an LSTM layer, and a fully connected layer; The first convolutional layer inputs the key features and outputs the first updated features; The second convolutional layer inputs the first updated features and outputs the second updated features; The LSTM layer inputs the second updated feature and outputs the third updated feature; The fully connected layer inputs the third updated feature and outputs the probability of occurrence of each fault type.

8. The avionics system fault prediction method according to claim 1, characterized in that: In step S5, the following steps are specifically included: S51, using the historical key features of the avionics system as training samples, making sample labels according to the corresponding fault labels, and constructing a data set according to the training samples and the sample labels; S52, constructing an original target prediction model according to the preliminary fault prediction model, wherein the architecture of the original target prediction model is the same as the architecture of the preliminary fault prediction model, and various parameters in the first convolution layer, the second convolution layer, and the LSTM layer are consistent; S53. Use the data set to train the original target prediction model to obtain a target prediction model.

Images