Fault Feature Extraction and Diagnosis Method for Aero-Converter Based on Multi-Source Information

Abstract

The present invention relates to the technical field of aviation equipment fault detection, and discloses a method for extracting fault features and diagnosing aviation converters based on multi-source information. The key points of the technical solution are to collect and preprocess time series data and spatial distribution data respectively to improve the accuracy and stability of the data. Then, time feature extraction, spatial feature extraction, spectrum acquisition and feature extraction are carried out on the preprocessed data respectively to realize the fusion of multi-source information; integrate information from different types of sensors and different feature domains, fully mine the multi-dimensional information of the equipment operation state, so as to improve the accuracy and reliability of fault detection. Finally, envelope detection is carried out on the signal to obtain the envelope line of the signal amplitude change, and the characteristic parameters of the envelope are analyzed to detect the fault information or pulse-type fault characteristics in the modulation signal; then, based on the deep convolutional neural network, the complex relationship between the envelope characteristics and the faults is learned to realize the diagnosis of aviation airborne converter faults.

Description

Technical Field

[0001] The present invention relates to the technical field of aviation equipment fault detection, and more specifically, it relates to a method for extracting fault characteristics and diagnosing faults of an aviation converter based on multi-source information. Background Art

[0002] Traditional fault detection methods for aviation airborne converters mainly rely on single signal analysis means, such as only performing time-domain analysis on electrical signals such as voltage and current, or only focusing on the temperature change of the equipment. These methods cannot comprehensively and accurately extract fault characteristic quantities. On the one hand, single signal analysis is easily affected by noise interference, resulting in the masking of fault characteristics; on the other hand, it is impossible to make full use of multi-dimensional information during the operation of the equipment, and the detection ability for some complex faults and early faults is insufficient. For example, when there is a slight local overheating fault inside the converter, it is very difficult to detect it in time only through the time-domain analysis of electrical signals, and the single-point measurement of the temperature sensor cannot accurately reflect the spatial distribution of the fault.

[0003] Therefore, the present invention provides a method for extracting fault characteristics and diagnosing faults of an aviation converter based on multi-source information, which improves the above technical problems. Summary of the Invention

[0004] The embodiments of the present disclosure aim at the deficiencies of the prior art and provide a method for extracting fault characteristics and diagnosing faults of an aviation converter based on multi-source information. The present invention uses multi-source information fusion technology to integrate information from different types of sensors and different feature domains, and fully excavates the multi-dimensional information of the equipment operation state, thereby improving the accuracy and reliability of fault detection.

[0005] The above technical object of the present invention is achieved through the following technical solutions: A method for extracting fault characteristics and diagnosing faults of an aviation converter based on multi-source information, including the following steps:

[0006] S1. Collect and preprocess time series data and spatial distribution data respectively;

[0007] S2. Perform time feature extraction, spatial feature extraction, spectrum acquisition and feature extraction on the preprocessed data respectively to achieve the fusion of multi-source information.

[0008] As a preferred technical solution of the present invention, the time series data collection process is: through sensors installed at key parts of the aviation airborne converter, continuously collect the data series of electrical parameters and environmental parameters during the operation of the converter to form a data set in the time dimension;

[0009] The process of spatial distribution data acquisition is as follows: Multiple sensors are arranged at different physical positions of the converter to obtain information data of different spatial regions inside the converter; these spatially distributed sensor data are organized and numbered according to the physical structure and circuit topology of the converter to establish the corresponding relationship between the data and the spatial positions, forming a data set in the spatial dimension.

[0010] As a preferred technical solution of the present invention, the steps of time series data preprocessing include: wavelet decomposition, noise standard deviation estimation, threshold determination, threshold processing, and wavelet reconstruction;

[0011] The steps of spatial distribution data preprocessing include: information entropy calculation, anomaly determination, and spatial interpolation.

[0012] As a preferred technical solution of the present invention, the process of wavelet decomposition is as follows: The collected discrete time series data is subjected to discrete wavelet transform and processed using the Coiflet wavelet Coif3 with good symmetry and regularity: The approximate coefficient of the -th layer decomposition is calculated by the formula: , where , is the low-pass filter coefficient corresponding to the selected wavelet basis, is its conjugate; the detail coefficient of the -th layer decomposition is calculated by the formula: , where , is the high-pass filter coefficient corresponding to the selected wavelet basis, is its conjugate;

[0013] The process of noise standard deviation estimation is as follows: Calculate the noise standard deviation estimation value of the detail coefficient ; adopt the median-based estimation method: ;

[0014] The process of threshold determination is as follows: The initial threshold is calculated by the formula: ;

[0015] The dynamic adjustment factor , where is the weight coefficient, is the temperature change amount, is the maximum allowable temperature change amount; the final threshold is ;

[0016] The process of threshold processing is as follows: Let be the mixing parameter , the processed detail coefficients are: when the time, ;

[0017] Among them, the soft threshold part can smooth the noise, and the hard threshold part can retain the sharp features of the signal;

[0018] The process of wavelet reconstruction is: using the processed detail coefficients and the approximation coefficients to perform inverse wavelet transform to reconstruct the denoised signal The formula is: , where is the maximum number of decomposition levels.

[0019] As a preferred technical solution of the present invention, the process of information entropy calculation is: for the data collected by the th sensor at a certain moment; first, calculate the mean value of the data in this neighborhood; then, calculate the information entropy , and the formula is , where ;

[0020] The process of anomaly determination is: set a normal information entropy threshold range for each sensor neighborhood; collect all the information entropy data of the sensor neighborhoods during the continuous normal operation of the converter for a period of time, and calculate its mean value and the standard deviation , set , , where and are empirical coefficients and can be adjusted according to the actual situation, such as ; when the information entropy of a certain sensor neighborhood exceeds this threshold range, that is, or , it is determined that there may be an anomaly in the spatial area where this neighborhood is located;

[0021] The process of spatial interpolation is: for any point in the converter space, the estimated value of its physical quantity is calculated by the following formula: , where is the measured value of the physical quantity at the known sensor position , is the number of sensors participating in the interpolation, is the weight coefficient, which is obtained by solving the following system of equations: , here is the semivariogram, which is used to describe the variability of physical quantities between two points in space and The Gaussian semivariogram , where is the distance between two points, is the nugget effect, is the sill value, is the range. These parameters are determined by the variogram analysis of sensor data.

[0022] As a preferred technical solution of the present invention, the process of time feature extraction is as follows: find all local maximum and minimum points of the time series , and obtain the upper envelope and the lower envelope respectively through cubic spline interpolation, calculate the mean value of the upper and lower envelopes, subtract the mean value from the original sequence to obtain ; if meets the IMF condition, then it is the first IMF component; otherwise, use as the new sequence and repeat the above process until a that meets the conditions is obtained; then, subtract from the original sequence to obtain the remaining sequence , and repeat the above decomposition steps for to sequentially obtain and the residual term , that is ;

[0023] To more accurately reflect the internal characteristics of the multi-variable time series of the aviation airborne converter, according to the importance of each parameter of the converter to the fault, a weight is assigned to each variable; the coarse-grained sequence is calculated as ;

[0024] For the coarse-grained sequence , calculate its sample entropy; given the embedding dimension and the tolerance , define the vector ; let represent the Chebyshev distance between two vectors and , that is ; define as the ratio of the number of that satisfy to ; then ; similarly, calculate ​of the ; sample entropy is ;

[0025] For the time series , its autocorrelation function , where is the mean of the sequence ;

[0026] For the time series , when calculating the autocorrelation, consider the subsequences at different delays and ; The DTW algorithm finds a path , where , such that and the distance between them is minimized; The constraints of the path are , , and ; The distance metric uses the Euclidean distance ; Define the DTW distance ; The improved autocorrelation function ;

[0027] For two time series and , the traditional cross-correlation function ; Similarly, the DTW algorithm is introduced for improvement; For the subsequences at different delays and , find the optimal path through the DTW algorithm, and calculate the DTW distance ; The improved cross-correlation function .

[0028] As a preferred technical solution of the present invention, the process of spatial feature extraction is as follows: Construct a deep convolutional neural network to extract the fused spatial features; The network structure includes multiple convolutional layers, pooling layers, and fully connected layers;

[0029] The convolutional layer: Perform convolutional operations on the input feature map using convolutional kernels of different sizes ; For the input feature map , the elements of the output feature map of the convolutional layer are calculated as , where are the elements of the convolutional kernel, is the bias, are the elements of the input feature map;

[0030] The pooling layer: uses max pooling. For the input feature map , within the pooling window , the elements of the output feature map are obtained, reducing the data dimension while retaining key features;

[0031] The fully connected layer: flattens the pooled feature map and performs a linear transformation through the weight matrix and the bias , and then passes through the activation function ReLU to obtain the final output; the output layer uses the Softmax function to map the result to the probability distribution of the fault categories , is the output of the fully connected layer, is the number of fault categories; the network is trained with a large amount of spatial feature data with fault labels, enabling the network to learn the complex relationship between spatial features and faults and achieving accurate fault location;

[0032] The method based on multi-physical field coupling and deep learning can deeply extract the spatial features of the aviation airborne converter, improving the accuracy and reliability of fault location.

[0033] As a preferred technical solution of the present invention, the process of spectrum acquisition is as follows: Define the Hilbert transform of each mode function , and combine it with to form the analytic signal , demodulate it to the baseband through the exponential term , and obtain ; take the square of the norm of its first derivative , that is , and use this as the bandwidth metric of each mode function; construct the constrained variational problem as , and the constraint condition is , where is the preset number of decomposed modes;

[0034] By introducing the quadratic penalty factor and the Lagrange multiplier , the constrained variational problem is transformed into an unconstrained augmented Lagrangian function ; use the alternating direction multiplier method to iteratively solve this augmented Lagrangian function, update and , until the convergence condition is satisfied;

[0035] After VMD decomposition, the original signal is decomposed into IMF components ​, each component has different frequency characteristics; VMD adaptively separates different frequency components, avoiding problems such as spectral aliasing, and provides a purer signal component for subsequent spectral feature extraction;

[0036] The process of spectral feature extraction is as follows: Time-frequency atom decomposition is used to decompose the signal into a linear combination of a series of atom functions with specific time-frequency characteristics; Gaussian Hermite time-frequency atom functions are selected , where is the normalization constant, is the Hermite polynomial of order represents the time center, represents the frequency center, controls the width of the time-frequency window;

[0037] For a given signal , the best time-frequency atom is found through the matching pursuit algorithm to approximate the signal; the specific process is as follows:

[0038] Initialize the residual ;

[0039] For each iteration Calculate the inner product of the residual with all possible time-frequency atoms , and find the time-frequency atom with the largest ;

[0040] Update the residual ;

[0041] After iterations, the signal can be approximately represented as ;

[0042] Based on the time-frequency atom decomposition results, recalculate the frequency centroid, bandwidth and other characteristics of the spectrum; the traditional frequency centroid formula is , in the framework of time-frequency atom decomposition, consider the contribution of each time-frequency atom to the frequency centroid; let , then the improved frequency centroid is ;

[0043] Introduce a bandwidth calculation method based on time-frequency atoms; first calculate the effective bandwidth of each time-frequency atom, which can be determined by the frequency support range or the relevant energy distribution of the time-frequency atom; then, the bandwidth is calculated as .

[0044] ​A fault diagnosis method for an aviation converter based on multi-source information includes the following steps: empirical mode decomposition, Hilbert transform to obtain the envelope, calculation of envelope characteristic parameters, construction of a deep convolutional neural network, model training and optimization.

[0045] As a preferred technical solution of the present invention, the process of empirical mode decomposition is as follows: find all the local maximum points and minimum points of the signal respectively obtain the upper envelope line and the lower envelope line by cubic spline interpolation, calculate the mean value of the upper and lower envelope lines , subtract the mean value from the original signal to obtain ; if satisfies condition, then it is the first component; otherwise, take as the new sequence and repeat the above process until the that meets the conditions is obtained; then, subtract from the original signal to obtain the remaining sequence , and repeat the above decomposition steps for to obtain and the residual term , that is, ;

[0046] The process of Hilbert transform to obtain the envelope is as follows: perform Hilbert transform on each to obtain ; thus construct the analytic signal , and its amplitude is the instantaneous amplitude of , that is, the envelope line;

[0047] The process of calculating envelope characteristic parameters is as follows: envelope mean value , where is the time length of the signal; the envelope mean value reflects the average level of the envelope amplitude and can be used to judge the overall offset of the signal amplitude;

[0048] Envelope variance , the envelope variance measures the degree of dispersion of the envelope amplitude around the mean value and can reflect the fluctuation size of the envelope amplitude;

[0049] Envelope peak factor , the envelope peak factor is used to evaluate the relative size of the peak value and the mean value in the envelope and is very sensitive to the detection of pulse-type faults. When a pulse-type fault occurs, the peak factor will increase significantly;

[0050] The process of constructing a deep convolutional neural network is as follows: the network structure includes multiple convolutional layers, pooling layers and fully connected layers;

[0051] Convolutional layer: In the convolutional layer, convolutional kernels of different sizes are used to perform convolutional operations on the input feature matrix; for the input feature matrix , the convolutional layer outputs a feature map elements The calculation formula is , where are the elements of the convolutional kernel, is the bias, are the elements of the input feature matrix; the convolutional layer can automatically extract local feature patterns in the feature matrix, different combinatorial relationships between envelope feature parameters;

[0052] Pooling layer: The maximum pooling method is used to reduce the dimensionality of the feature map output by the convolutional layer; within the pooling window of size , the index of the output feature map; here, the pooling layer reduces the data volume, reduces the computational complexity, and enhances the translational invariance of the model to the input without losing key features;

[0053] Fully connected layer: The feature map processed by the pooling layer is flattened and connected to the fully connected layer; the output of the fully connected layer is linearly transformed through the weight matrix and the bias to obtain the final fault diagnosis result after passing through the Softmax function; for the input vector , the output of the fully connected layer is , where function maps the output value to the interval, and the sum of all output values is 1, and each output value represents the probability of the corresponding fault category;

[0054] The process of model training and optimization is as follows: Use the envelope feature data of the aviation airborne converter signal with fault labels to train the DCNN; during the training process, update the weight parameters of the network through the cross-entropy loss function , where is the number of samples, is the number of fault categories, is the sample belongs to the category true label, is the probability that the model predicts that the sample belongs to the category ; at the same time, use stochastic gradient descent to accelerate model convergence and improve training efficiency.

[0055] In summary, the present invention has the following beneficial effects:

[0056] First, collect time series data and spatial distribution data through sensors respectively to form a time - dimension data set and a spatial - dimension data set, and pre - process the data sets in the two dimensions to improve the accuracy and stability of the data.

[0057] Second, through time - feature extraction, the features of the time series of the aircraft - borne converter can be extracted more deeply and accurately. Through spatial - feature extraction, the spatial features of the aircraft - borne converter can be deeply extracted to improve the accuracy and reliability of fault location. Through spectrum acquisition and feature extraction, the frequency components and energy distribution of the signal are determined. Based on the above time - space - spectrum feature extraction, the fusion of multi - source information is realized, which can integrate information from different types of sensors and different feature domains, fully mine the multi - dimensional information of the equipment operation state, and thus improve the accuracy and reliability of fault detection.

[0058] Third, first perform envelope detection on the signal to obtain the envelope line of the signal amplitude change, and analyze the characteristic parameters of the envelope to detect fault information or pulse - type fault characteristics in the modulated signal; then, based on the deep convolutional neural network, learn the complex relationship between the envelope characteristics and faults to realize the diagnosis of faults in the aircraft - borne converter. Brief Description of the Drawings

[0059] Figure 1 It is a flowchart of the fault feature extraction of the aircraft converter based on multi - source information provided by the embodiment of the present invention.

[0060] Figure 2 It is a flowchart of the fault diagnosis method of the aircraft converter based on multi - source information provided by the embodiment of the present invention. Detailed Embodiments

[0061] The following will explain the present application in detail with specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that those of ordinary skill in the art can make several deformations and improvements without departing from the concept of the present application. These all belong to the protection scope of the present application.

[0062] In order to make the purpose, technical solution and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application.

[0063] It should be noted that if there is no conflict, the various features in the embodiments of the present application can be combined with each other, and all are within the protection scope of the present application. In addition, although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order from the module division in the device or the flowchart. In addition, the terms "first", "second", "third", etc. used herein do not limit the data and the execution order, but only distinguish the same items or similar items with basically the same functions and effects.

[0064] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs. The terms used in this specification in the description of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. The term "and / or" used in this specification includes any and all combinations of one or more of the related listed items.

[0065] In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0066] The embodiments of the present disclosure aim to solve the problem that the traditional fault detection method for aviation airborne converters relies on a single signal analysis method and cannot comprehensively and accurately extract fault feature quantities. In view of this, the embodiments of the present disclosure propose a method for extracting and diagnosing fault features of aviation converters based on multi-source information. Through multi-source information fusion, information from different types of sensors and different feature domains can be integrated, and multi-dimensional information on the operating state of the equipment can be fully mined, thereby improving the accuracy and reliability of fault detection.

[0067] Please refer to Figure 1 , Figure 1 which shows the flowchart of the method for extracting and diagnosing fault features of aviation converters based on multi-source information according to the embodiments of the present disclosure. The overall process mainly includes the following 5 steps:

[0068] Step 1: Time series data acquisition and preprocessing.

[0069] Specifically, through sensors installed at key parts of the aviation airborne converter, a data sequence of electrical parameters and environmental parameters during the operation of the converter is continuously collected to form a data set in the time dimension.

[0070] The collected time series data is denoised, and a filtering algorithm is used to remove high-frequency noise and measurement errors to improve the accuracy and stability of the data.

[0071] Example 1. The steps of time series data acquisition and preprocessing include: S1.1 Wavelet decomposition, S1.2 Noise standard deviation estimation, S1.3 Threshold determination, S1.4 Threshold processing, and S1.5 Wavelet reconstruction.

[0072] S1.1. For the collected discrete time series data , where N is the data length, perform discrete wavelet transform and use the Coiflet wavelet Coif3 with good symmetry and regularity for processing.

[0073] The approximate coefficient of the layer decomposition is calculated as: , where , is the low-pass filter coefficient corresponding to the selected wavelet basis, is its conjugate. This step filters the signal to extract the low-frequency components of the signal, reflecting the general trend of the signal. In the current data of the aviation converter, the low-frequency components usually represent the basic current level during the normal operation of the converter.

[0074] The detail coefficient of the layer decomposition is calculated as: , where , is the high-pass filter coefficient corresponding to the selected wavelet basis, is its conjugate. This step extracts the high-frequency components of the signal, and the high-frequency noise is concentrated in these detail coefficients. For example, the high-frequency noise such as electromagnetic interference received by the aviation converter is reflected in the detail coefficients.

[0075] S1.2. Calculate the noise standard deviation estimation value of the detail coefficient . Adopt the median-based estimation method: . In the aviation environment, the noise usually has non-Gaussian and sudden characteristics. This median-based estimation method is more robust than the traditional mean-based estimation method and can more accurately estimate the noise intensity.

[0076] S1.3. Considering the dynamic changes in the operating state of the aviation airborne converter, combine factors such as the temperature change rate of the converter to adjust the threshold. The threshold determination method adopted is:

[0077] The initial threshold is calculated as: .

[0078] The dynamic adjustment factor , where is the weight coefficient, is the temperature change amount, is the maximum allowable temperature change amount.

[0079] Final threshold is . Through the above steps, the threshold can be adaptively adjusted according to the actual operating state of the converter, and can more effectively remove noise when the noise intensity changes or the operating conditions of the converter change.

[0080] S1.4. Adopt an improved method combining soft and hard thresholds. Let be the mixing parameter , and the processed detail coefficient is:

[0081] When , .

[0082] Here, the soft threshold part can smooth the noise, and the hard threshold part can retain the sharp features of the signal.

[0083] When , .

[0084] S1.5. Use the processed detail coefficient and the approximation coefficient to perform wavelet inverse transform to reconstruct the denoised signal The formula is: , where is the maximum decomposition level. Through wavelet inverse transform, the denoised high-frequency and low-frequency components are recombined to obtain the denoised signal, improving the accuracy and stability of the data and providing a reliable data basis for subsequent fault detection.

[0085] Step 2: Spatial distribution data acquisition and preprocessing.

[0086] Specifically, multiple sensors are arranged at different physical positions of the converter to obtain information on different spatial regions inside the converter. Temperature sensors are set at positions such as power modules, heat sinks, and control circuit boards to monitor the spatial distribution of temperature; voltage and current sensors are set at the input and output ends and key circuit nodes to obtain the spatial distribution information of electrical parameters. Organize and number these spatially distributed sensor data according to the physical structure and circuit topology of the converter, establish the correspondence between the data and the spatial positions, and form a data set in the spatial dimension.

[0087] Example 2. The steps of spatial distribution data acquisition and preprocessing include: S2.1 Information entropy calculation, S2.2 Anomaly determination, S2.3 Spatial interpolation.

[0088] S2.1. For the data collected by each sensor, consider the information entropy of its spatial neighborhood. Assume that in the converter space, with a certain sensor as the center, its neighborhood is defined as the set of the sensor and its directly adjacent surrounding sensors. Suppose there are sensors in this neighborhood. For the data collected by the -th sensor at a certain moment.

[0089] First, calculate the mean value of the data in this neighborhood. Then, calculate the information entropy using the formula , where . The information entropy reflects the uncertainty or chaos degree of the data in the neighborhood. When the data in the neighborhood varies greatly, the information entropy is high; while when the data is relatively consistent, the information entropy is low. In a certain area of the heat sink, if the measured values of the temperature sensors vary greatly, it indicates that there may be an abnormality in the heat dissipation of this area, and at this time the information entropy will be relatively high.

[0090] S2.2. Set a normal information entropy threshold range for each sensor neighborhood. This threshold range is obtained through statistical analysis of a large number of spatial neighborhood information entropy data of the converter in the normal operating state. Collect all the spatial neighborhood information entropy data of the sensors during a period of continuous normal operation of the converter, and calculate its mean value and standard deviation . Set , , where and are empirical coefficients, which can be adjusted according to the actual situation, such as .

[0091] When the information entropy of a certain sensor neighborhood exceeds this threshold range, that is, or , it is determined that the spatial area where this neighborhood is located may have an abnormality. If the information entropy of the temperature sensor neighborhood near a certain power module suddenly increases and exceeds the normal range, it may mean that there are abnormal situations such as local overheating or uneven heat dissipation in this power module.

[0092] S2.3. Since the distribution of sensors in the converter space is discrete, in order to more comprehensively understand the variation of physical quantities in space, the Kriging interpolation method is adopted.

[0093] For any point in the converter space, the estimated value of its physical quantity is calculated by the following formula: , where is the measured value of the physical quantity at the known sensor position , is the number of sensors participating in interpolation is the weight coefficient

[0094] is obtained by solving the following system of equations , where is the semivariogram, which is used to describe the variability of the physical quantity between two points in space and . The Gaussian semivariogram , where is the distance between two points is the nugget effect is the sill value is the range. These parameters are determined by analyzing the variogram of the sensor data. Through spatial interpolation, a continuous distribution of the physical quantity in the converter space is obtained, which can more accurately capture the change trend of the physical quantity

[0095] Step 3: Time feature extraction

[0096] The statistical features of time series data can reflect the change trend, fluctuation degree and stability of the signal over time. Extracting the trend features of the time series to determine the long-term trend of the signal and judge whether there is an abnormal upward or downward trend, which helps to detect the slow change of the converter performance

[0097] Example 3: The specific algorithm is as follows

[0098] Empirical Mode Decomposition (EMD) is an adaptive decomposition method for non-linear and non-stationary time series. It decomposes the time series into several Intrinsic Mode Functions (IMFs) and a residue . Each IMF needs to meet two conditions: one is that within the entire data segment, the number of extreme points (maximum and minimum values) must be equal to or at most differ by one from the number of zero-crossing points; the other is that at any moment, the mean value of the upper and lower envelope lines formed by the local maximum points and local minimum points respectively is zero

[0099] First, find all the local maximum points and minimum points of the time series , and respectively obtain the upper envelope line and the lower envelope line through cubic spline interpolation, calculate the mean value of the upper and lower envelope lines, and subtract the mean value from the original sequence to obtain . If meets the IMF conditions, then it is the first IMF component; otherwise, Repeat the above process with the new sequence until the condition is met. Then, subtract from the original sequence to obtain the remaining sequence . For , repeat the above decomposition steps to successively obtain and the residual term , that is .

[0100] In the time series data of an aircraft airborne converter, EMD can separate fluctuations on different time scales. The high-frequency IMF components may reflect short-term disturbances or noises during the operation of the converter, while the low-frequency IMF components and the residual term contain more trend information related to progressive faults.

[0101] By analyzing the autocorrelation function and cross-correlation function of the time series, determine the correlation of the signal at different time delays, and detect the periodic fault characteristics and the mutual relationship between signals. The specific algorithm is as follows:

[0102] To more accurately reflect the internal characteristics of the multivariate time series of an aircraft airborne converter, assign weights to each variable according to the importance of the converter parameters to the fault influence . The coarse-grained sequence is calculated as .

[0103] For the coarse-grained sequence , calculate its sample entropy. Given the embedding dimension and the tolerance , define the vector . Let represent the Chebyshev distance between two vectors and , that is . Define as the ratio of the number of that satisfy to . Then . Similarly, calculate the of the -dimensional vector. The sample entropy is . By this method, more accurately characterize the complexity of the multivariate time series of the converter at different scales, and provide more effective features for fault diagnosis.

[0104] For the time series , its autocorrelation function , where is the mean value of the sequence .​

[0105] However, the time series of the aircraft-borne converter may have local deformations or phase differences, which affect the accuracy of autocorrelation analysis.

[0106] Therefore, an autocorrelation improvement using dynamic time warping (DTW) is introduced, and the DTW algorithm is used to solve this problem. For time series , when calculating the autocorrelation, subsequences at different delays and are considered. The DTW algorithm finds a path , where , such that the distance between and is minimized. The constraints for the path are , , and . The distance metric uses the Euclidean distance . The DTW distance is defined as . The improved autocorrelation function is . This can more accurately capture the similarity of the time series at different delays, and can effectively analyze it even in the presence of local deformations, which helps to discover periodic fault characteristics in the converter time series.

[0107] For two time series and , the traditional cross-correlation function is . The DTW algorithm is also introduced for improvement. For subsequences at different delays and , the optimal path is found through the DTW algorithm, and the DTW distance is calculated. The improved cross-correlation function is . This improved cross-correlation analysis can more precisely determine the mutual relationship between different signals, providing stronger support for the fault diagnosis of aircraft-borne converters based on multivariate time series analysis.

[0108] Through the above improved multivariate multi-scale sample entropy and the autocorrelation and cross-correlation analysis methods combined with DTW, the characteristics of the aircraft-borne converter time series can be extracted more deeply and accurately.

[0109] Step 4: Spatial feature extraction.

[0110] A deep convolutional neural network (DCNN) is constructed to extract the fused spatial features.

[0111] Embodiment 4: The network structure includes: multiple convolutional layers, pooling layers and fully connected layers.

[0112] Convolutional layer: using convolution kernels of different sizes Perform convolution operation on the input feature map. , the convolutional layer outputs feature maps Elements Calculated as ,in is the convolution kernel element, is the bias, It is the input feature map element. The convolution operation extracts local spatial features, such as local temperature change patterns, thermal stress concentration areas, etc.

[0113] Pooling layer: using maximum pooling, for the input feature map , in the pooling window Inside, output feature map Elements , reducing the data dimension while retaining key features.

[0114] Fully connected layer: Flatten the pooled feature map and pass it through the weight matrix and bias Perform linear transformation and then pass the activation function ReLU to get the final output. The output layer uses the Softmax function to map the result to the probability distribution of the fault category. , is the output of the fully connected layer, is the number of fault categories. The network is trained by a large amount of spatial feature data with fault labels, so that the network can learn the complex relationship between spatial features and faults and achieve accurate fault location.

[0115] Through this method based on multi-physics field coupling and deep learning, the spatial characteristics of aviation airborne inverters can be deeply extracted, improving the accuracy and reliability of fault location.

[0116] Step 5: Spectrum acquisition and feature extraction.

[0117] Perform short-time Fourier transform (STFT) on the collected time series data (especially current and voltage signals), convert the time domain signal into a frequency domain signal, obtain the signal's spectrum data and characteristics, calculate the amplitude and phase of each harmonic, and determine the signal's frequency component and energy distribution.

[0118] Embodiment 5, the specific algorithm is as follows:

[0119] Variational mode decomposition (VMD) is an adaptive signal decomposition method that aims to decompose complex signals into a series of intrinsic mode functions (IMFs) with different center frequencies. For a given input signal , The goal of VMD is to determine a series of modal functions by solving a variational problem and the central frequency .

[0120] Define the Hilbert transform of each modal function , combine it with to form an analytic signal , demodulate it to the baseband through the exponential term to obtain . Take the square of the norm of its first derivative , that is , and use this as the bandwidth metric for each modal function. Construct the constrained variational problem as , and the constraint condition is , where is the preset number of decomposition modes .

[0121] By introducing the quadratic penalty factor and the Lagrange multiplier , transform the constrained variational problem into an unconstrained augmented Lagrangian function . Use the alternating direction method of multipliers (ADMM) to iteratively solve this augmented Lagrangian function, update and until the convergence condition is satisfied

[0122] After VMD decomposition, the original signal is decomposed into IMF components , and each component has different frequency characteristics. VMD adaptively separates different frequency components, avoiding problems such as spectral aliasing, and provides a cleaner signal component for subsequent spectral feature extraction

[0123] The purpose of using time-frequency atom decomposition is to decompose the signal into a linear combination of a series of atom functions with specific time-frequency characteristics. Select the Gaussian Hermite time-frequency atom function , where is the normalization constant is the Hermite polynomial of order represents the time center represents the frequency center controls the width of the time-frequency window

[0124] For a given signal , find the best time-frequency atom to approximate the signal through the matching pursuit algorithm. The specific process is as follows

[0125] Initialize the residual .

[0126] For each iteration Calculate the residual with all possible time-frequency atoms for the inner product and find the largest time-frequency atom .

[0127] Update the residual .

[0128] After iterations, the signal can be approximately represented as .

[0129] Based on the time-frequency atom decomposition results, recalculate the frequency centroid and bandwidth and other characteristics of the spectrum. The traditional frequency centroid formula is , and in the time-frequency atom decomposition framework, consider the contribution of each time-frequency atom to the frequency centroid. Let , then the improved frequency centroid is . This calculation method can better reflect the frequency contribution of the signal in different time-frequency localities. For the converter signal, it can more accurately capture the frequency centroid shift caused by component failures.

[0130] Introduce a bandwidth calculation method based on time-frequency atoms. First, calculate the effective bandwidth of each time-frequency atom, which can be determined by the frequency support range or related energy distribution of the time-frequency atom. Then, the bandwidth is calculated as . This formula comprehensively considers the frequency position, amplitude of each time-frequency atom and its relationship with the frequency centroid, and can more accurately measure the discrete degree of the signal frequency distribution, and is more sensitive to detecting the bandwidth change caused by converter faults.

[0131] Please refer to Figure 2 , Figure 2 which shows the flowchart of the aviation converter fault diagnosis method based on multi-source information described in the embodiments of the present disclosure.

[0132] By performing envelope detection on the signal, obtain the envelope of the signal amplitude change, and analyze the characteristic parameters of the envelope to detect the fault information or pulse-type fault characteristics in the modulation signal.

[0133] Embodiment Six: The overall process mainly includes the following 5 steps: empirical mode decomposition (EMD), Hilbert transform to obtain the envelope, envelope characteristic parameter calculation, construction of a deep convolutional neural network (DCNN), model training and optimization.

[0134] S1. Use empirical mode decomposition (EMD) to decompose the complex non-stationary signal into multiple intrinsic mode functions (IMFs) and a residue . Each needs to meet two conditions: one is that within the entire data segment, the number of extreme points (maxima and minima) must be equal to or at most differ by one from the number of zero-crossing points; the other is that at any given time, the mean of the upper and lower envelopes formed by the local maximum points and local minimum points respectively is zero. The specific decomposition process is as follows:

[0135] Find all the local maximum and minimum points of the signal , and obtain the upper envelope and the lower envelope respectively through cubic spline interpolation. Calculate the mean of the upper and lower envelopes , and subtract the mean from the original signal to obtain . If meets the condition, then it is the first component; otherwise, use as the new sequence and repeat the above process until a that meets the condition is obtained. Then, subtract from the original signal to obtain the remaining sequence . Repeat the above decomposition steps for to successively obtain and the residue , that is .

[0136] S2. Perform Hilbert transform on each to obtain . Thus, construct the analytic signal , whose amplitude is the instantaneous amplitude of , that is, the envelope. In this way, the envelope characteristics of the signal at different time scales can be accurately extracted. These envelope characteristics reflect the change of the signal amplitude and are of great significance for detecting fault information in modulated signals or pulse-type fault characteristics. In the current signal of an aircraft-borne converter, if there is a pulse-type fault, its corresponding envelope will show specific mutations or abnormal fluctuations.

[0137] S3. To more comprehensively describe the envelope characteristics, calculate the following characteristic parameters:

[0138] Envelope mean , where is the time length of the signal. The envelope mean reflects the average level of the envelope amplitude and can be used to judge the overall offset of the signal amplitude.

[0139] Envelope variance , and the envelope variance measures the degree of dispersion of the envelope amplitude around the mean, which can reflect the fluctuation magnitude of the envelope amplitude.

[0140] Envelope peak factor , and the envelope peak factor is used to evaluate the relative size of the peak value and the mean value in the envelope. It is very sensitive to the detection of pulse-type faults. When a pulse-type fault occurs, the peak factor will increase significantly.

[0141] S4. Compose the extracted envelope feature parameters (envelope mean, envelope variance, envelope peak factor, etc.) into a feature matrix as the input of the deep convolutional neural network. The network structure includes multiple convolutional layers, pooling layers, and fully connected layers.

[0142] Convolutional layer: In the convolutional layer, convolutional kernels of different sizes are used to perform convolution operations on the input feature matrix. For the input feature matrix , the convolutional layer outputs a feature map whose elements are calculated by the formula , where are the elements of the convolutional kernel, is the bias, are the elements of the input feature matrix. The convolutional layer can automatically extract the local feature patterns in the feature matrix and the combined relationships between different envelope feature parameters.

[0143] Pooling layer: The maximum pooling method is used to reduce the dimension of the feature map output by the convolutional layer. Within the pooling window of size , the indices of the output feature map are obtained. Here, the pooling layer reduces the data volume and computational complexity without losing key features, and at the same time enhances the translational invariance of the model to the input.

[0144] Fully connected layer: Flatten the feature map processed by the pooling layer and connect it to the fully connected layer. The output of the fully connected layer is linearly transformed through the weight matrix and the bias , and then the final fault diagnosis result is obtained through the Softmax function. For the input vector , the output of the fully connected layer is , where the function maps the output value to the interval, and the sum of all output values is 1. Each output value represents the probability of the corresponding fault category.

[0145] S5. Use the envelope feature data of the aerospace airborne converter signals with fault labels to train the DCNN. During the training process, update the weight parameters of the network through the cross-entropy loss function where is the number of samples, is the number of fault categories, is the sample belonging to the category true label, is the probability that the model predicts the sample belongs to the category . Meanwhile, adopt stochastic gradient descent (SGD) to accelerate the model convergence and improve the training efficiency.

[0146] In this way, the deep convolutional neural network can learn the complex relationship between the envelope features and faults, and realize the diagnosis of aerospace airborne converter faults.

[0147] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for extracting fault features of aviation inverters based on multi-source information, characterized in that: The method comprises the following steps: S1, collect and preprocess time series data and spatial distribution data respectively; S2, respectively performing time feature extraction, spatial feature extraction, spectrum acquisition and feature extraction on the pre-processed data to achieve fusion of multi-source information; The steps of time series data preprocessing include: wavelet decomposition, noise standard deviation estimation, threshold determination, threshold processing, and wavelet reconstruction; The steps of spatial distribution data preprocessing include: information entropy calculation, anomaly determination, and spatial interpolation; The process of temporal feature extraction is to find all local maximum and minimum points of the time series x(t), and obtain the upper envelope e by cubic spline interpolation. max (t) and the lower envelope e min (t), calculate the mean of the upper and lower envelopes Subtract the mean from the original sequence to obtain h1(t)=x(t)-m1(t); if h1(t) satisfies the IMF condition, it is the first IMF component; otherwise, take h1(t) as the new sequence and repeat the above process until the IMF1(t) that satisfies the condition is obtained; then, subtract IMF1(t) from the original sequence to obtain the remaining sequence r1(t)=x(t)-IMF1(t), repeat the decomposition steps for r1(t) to obtain IMF2(t), IMF3(t),…,IMF n (t) and the residual term r n (t), that is In order to more accurately reflect the intrinsic characteristics of the multivariate time series of aircraft airborne converters, a weight w is assigned to each variable according to the importance of each converter parameter on the fault. i , Coarse-grained sequence z j Calculated as For the coarse-grained sequence z j , calculate its sample entropy; given the embedding dimension n and tolerance r, define the vector make Represents two vectors and The Chebyshev distance of definition To satisfy The number of k and The ratio of k≠j; then Similarly, we calculate C for the n+1-dimensional vector n+1 (r); sample entropy SampEn(n,r,τ) is For the time series x(t), its autocorrelation function in is the mean of the sequence x(t); For a time series x(t), when calculating the autocorrelation, consider the subsequences x1 = {x(1), x(2), …, x(N-τ)} and x2 = {x(τ+1), x(τ+2), …, x(N)} under different delays τ; the DTW algorithm finds a path P = (p1, p2, …, p L ), where p l =(i l ,j l ), so that the distance between x1 and x2 is minimized; the constraints of path P are i1=1,j1=1,i L =N-τ,j L =N, and i l+1 ≥i l ,j l+1 ≥j l ; The distance metric uses Euclidean distance Defining DTW distance Improved autocorrelation function For two time series x(t) and y(t), the traditional cross-correlation function is The DTW algorithm is also introduced for improvement; for subsequences x1 = {x(1), x(2), ..., x(N-τ)} and y2 = {y(τ+1), y(τ+2), ..., y(N)} under different delays τ, the optimal path P is found through the DTW algorithm, and the DTW distance DTW(x1, y2) is calculated; the improved cross-correlation function The process of spatial feature extraction is: constructing a deep convolutional neural network to extract the fused spatial features; the network structure includes multiple convolutional layers, pooling layers and fully connected layers; The convolution layer: uses convolution kernels W of different sizes to perform convolution operations on the input feature map; for the input feature map F, the convolution layer outputs the element g of the feature map G i,j,k Calculated as g i,j,k =∑ m,n,l w m,n,l f i+m-1,j+n-1,k+l-1 +b, where w m,n,l is the convolution kernel element, b is the bias, f i,j,k is the input feature map element; The pooling layer: uses maximum pooling. For the input feature map G, within the pooling window s×s, the element h of the output feature map H i,j,k =max m,n g i×s+m,j×s+n,k , reducing data dimensionality while retaining key features; The fully connected layer: flattens the pooled feature map and passes the weight matrix W f and bias b f Perform linear transformation, and then use the activation function ReLU to get the final output; the output layer uses the Softmax function to map the result to the probability distribution of the fault category z i is the output of the fully connected layer, and C is the number of fault categories. The network is trained by a large amount of spatial feature data with fault labels, so that the network can learn the complex relationship between spatial features and faults and achieve accurate fault location. The method based on multi-physics field coupling and deep learning can deeply extract the spatial characteristics of aircraft-borne converters and improve the accuracy and reliability of fault location; The process of spectrum acquisition is: define each modal function u k The Hilbert transform H[u k (t)], and compare it with u k (t) constitutes the analytical signal u k (t)+jH[u k (t)], through the exponential term Demodulate to baseband and get L, whose first-order derivative is taken 2 The square of the norm, that is This is used as the bandwidth measure of each mode function; the constrained variation problem is constructed as The constraints are Where K is the preset number of decomposition modes; By introducing the quadratic penalty factor α and the Lagrangian multiplier λ(t), the constrained variational problem is transformed into an unconstrained augmented Lagrangian function: The augmented Lagrangian function is solved iteratively using the alternating direction multiplier method, and the update and λ n+1 , until the convergence condition is met; After VMD decomposition, the original signal f(t) is decomposed into K IMF components u k (t), each component has different frequency characteristics; VMD adaptively separates different frequency components, avoiding problems such as spectrum aliasing, and providing purer signal components for subsequent spectrum feature extraction; The process of spectrum feature extraction is as follows: time-frequency atomic decomposition is used to decompose the signal into a linear combination of a series of atomic functions with specific time-frequency characteristics; Gauss-Hermitian time-frequency atomic function is selected Among them A n,k is the normalization constant, H n (x) is an n-order Hermitian polynomial, τ represents the time center, ω represents the frequency center, and σ controls the width of the time-frequency window; For a given signal x(t), the matching pursuit algorithm is used to find the best time-frequency atom to approximate the signal; the specific process is as follows: Initialize residual r0(t)=x(t); For each iteration m, calculate the residual r m-1 (t) with all possible time-frequency atoms The inner product of turn up The largest time-frequency atom Update residual After M iterations, the signal x(t) can be approximately expressed as Based on the time-frequency atomic decomposition results, the frequency center of gravity and bandwidth of the spectrum are recalculated; the traditional frequency center of gravity formula is In the framework of time-frequency atomic decomposition, the contribution of each time-frequency atom to the frequency center of gravity is considered; let The improved frequency center of gravity for Introduce a bandwidth calculation method based on time-frequency atoms; first calculate the effective bandwidth Δω of each time-frequency atom m , which can be determined by the frequency support range or the relevant energy distribution of the time-frequency atom; then, the bandwidth B is calculated as 2. The method for extracting fault features of an aviation inverter based on multi-source information according to claim 1, characterized in that: The time series data collection process is as follows: through sensors installed at key locations of aircraft-borne converters, the data series of electrical parameters and environmental parameters of the converters during operation are continuously collected to form a data set in the time dimension; The spatial distribution data acquisition process is as follows: multiple sensors are arranged at different physical locations of the converter to obtain information data of different spatial regions inside the converter; These spatially distributed sensor data are organized and numbered according to the physical structure and circuit topology of the converter, and a corresponding relationship between the data and the spatial position is established to form a data set of spatial dimension.

3. The method for extracting fault features of an aviation inverter based on multi-source information according to claim 1, characterized in that: The process of wavelet decomposition is: perform discrete wavelet transform on the collected discrete time series data x(n), n = 0, 1, ..., N-1, and use Coiflet wavelet Coif3 with good symmetry and regularity for processing: the approximate coefficient c of the j-th layer decomposition j,k The calculation formula is: in, h j,k (n) is the low-pass filter coefficient corresponding to the selected wavelet basis, Its conjugate; the detail coefficient d of the j-th layer decomposition j,k The calculation formula is: in g j,k (n) is the high-pass filter coefficient corresponding to the selected wavelet basis, for its conjugation; The process of estimating the noise standard deviation is as follows: Calculate the detail coefficient d 1,k ,k=0,1,…,2 1 -1 noise standard deviation estimate Using a median-based estimation method: The process of determining the threshold is as follows: The calculation formula of the initial threshold λ0 is: Dynamic adjustment factor α: Where β1 is the weight coefficient, ΔT(t) is the temperature change, T max is the maximum allowable temperature change; the final threshold λ is: λ=αλ0; The threshold processing process is as follows: let γ be the mixing parameter 0<γ<1, and the processed detail coefficient For: When |d j,k When |>λ, Among them, the soft threshold part γ[sgn(d j,k )(|d j,k |-λ)] can smooth the noise, and the hard threshold part (1-γ)d j,k Can preserve the sharp features of the signal; The process of wavelet reconstruction is: using the processed detail coefficients and the approximation coefficient c j,k Perform inverse wavelet transform to reconstruct the denoised signal The formula is: Where J is the maximum number of decomposition levels.

4. The method for extracting fault features of an aviation inverter based on multi-source information according to claim 1, characterized in that: The process of calculating information entropy is as follows: for the data x collected by the i-th sensor at a certain moment i ; First, calculate x i The mean of the data in the neighborhood Then, calculate the information entropy H, the formula is in The process of abnormal judgment is: set a normal information entropy threshold range [H min ,H max ]; Collect all sensor neighborhood information entropy data when the converter operates normally for a period of time, calculate its mean μ and standard deviation σ, and set H min =μ-k1σ,H max =μ+k2σ, where k1 and k2 are empirical coefficients, k1=k2=1.5; when the information entropy H of a sensor neighborhood exceeds this threshold range, that is, H <H min or H>H max , then it is determined that there may be anomalies in the spatial area where the neighborhood is located; The process of spatial interpolation is: for any point P(x,y,z) in the converter space, the estimated value of the physical quantity u(P) is Calculated by the following formula: where u(x i ,y i ,z i ) is the known sensor position (x i ,y i ,z i ), m is the number of sensors involved in the interpolation, λ i is the weight coefficient, obtained by solving the following system of equations: Here γ(x i ,x j ) is the semivariogram function, which is used to describe the relationship between two points (x i ,y i ,z i ) and (x j ,y j ,z j ) between physical quantities, Gaussian semivariogram function in is the distance between two points, C0 is the nugget effect, C1 is the base value, and a is the range. These parameters are determined by analyzing the variation function of the sensor data.

5. The method for diagnosing faults of aviation inverters based on multi-source information is characterized in that: The method uses the aviation inverter fault feature extraction method based on multi-source information as described in any one of claims 1-4 for diagnosis, and includes the following steps: empirical mode decomposition, Hilbert transform envelope, envelope feature parameter calculation, construction of deep convolutional neural network, model training and optimization.

6. The aviation inverter fault diagnosis method based on multi-source information according to claim 5, characterized in that: The process of empirical mode decomposition is to find all local maximum and minimum points of the signal x(t), and obtain the upper envelope e by cubic spline interpolation. max (t) and the lower envelope e min (t), calculate the mean of the upper and lower envelopes Subtract the mean from the original signal to obtain h1(t)=x(t)-m1(t); if h1(t) satisfies the IMF condition, it is the first IMF component; otherwise, take h1(t) as a new sequence and repeat the above process until an IMF1(t) that satisfies the condition is obtained; then, subtract IMF1(t) from the original signal to obtain the remaining sequence r1(t)=x(t)-IMF1(t), repeat the above decomposition steps for r1(t), and obtain IMF2(t), IMF3(t),…, IMF n (t) and the residual term r n (t), that is The process of finding the envelope by Hilbert transform is: for each IMF i (t) is subjected to Hilbert transform, and we get The analytical signal z is constructed from this i (t) = IMF i (t)+jy i (t), its amplitude IMF i The instantaneous amplitude of , that is, the envelope; The process of calculating the envelope characteristic parameters is: envelope mean Where T is the time length of the signal; the envelope mean reflects the average level of the envelope amplitude and can be used to determine the overall deviation of the signal amplitude; Envelope Variance The envelope variance measures the dispersion of the envelope amplitude around the mean, and can reflect the fluctuation of the envelope amplitude; Envelope Crest Factor The envelope crest factor is used to evaluate the relative size of the peak value to the mean value in the envelope. It is very sensitive to detecting pulse-type faults. When a pulse-type fault occurs, the crest factor will increase significantly. The process of building a deep convolutional neural network is as follows: the network structure includes multiple convolutional layers, pooling layers and fully connected layers; Convolutional layer: In the convolutional layer, convolution kernels W of different sizes are used to perform convolution operations on the input feature matrix; for the input feature matrix X, the convolutional layer outputs the element y of the feature map Y i,j,k The calculation formula is y i,j,k =∑ m,n,l w m,n, l x i+m-1,j+n-1,k+l-1 +b, where w m,n,l is the convolution kernel element, b is the bias, x i,j,k is the input feature matrix element; The convolutional layer can automatically extract the local feature patterns in the feature matrix and the combination relationship between different IMF envelope feature parameters; Pooling layer: The maximum pooling method is used to reduce the dimension of the feature map output by the convolution layer; the index of the feature map is output within the pooling window of size s×s; here, the pooling layer reduces the amount of data and the computational complexity without losing key features, while enhancing the model's translation invariance to the input; Fully connected layer: Flatten the feature map processed by the pooling layer and connect it to the fully connected layer; the output of the fully connected layer is passed through the weight matrix W f and bias b f Linear transformation is performed, and then the final fault diagnosis result is obtained through the Softmax function; for the input vector V, the output p of the fully connected layer is p = Softmax (W f V+b f ), where the Softmax function maps the output value to the interval [0,1], and the sum of all output values ​​is 1, and each output value represents the probability of the corresponding fault category; The process of model training and optimization is as follows: DCNN is trained using the envelope feature data of aircraft airborne converter signals with fault labels; during the training process, the cross entropy loss function is used to To update the network weight parameters, where N is the number of samples, C is the number of fault categories, and y ij is the true label of sample i belonging to category j, p ij is the probability that the model predicts that sample i belongs to category j; at the same time, stochastic gradient descent is used to accelerate model convergence and improve training efficiency.

Images