Spacecraft high-dimensional data feature screening method based on multivariate feature extraction
By combining multi-domain physical feature extraction and stacked autoencoders, correlation analysis is used to screen performance degradation features from spacecraft telemetry data, solving the problem of difficulty in extracting effective features in existing technologies and achieving more accurate health management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-04-14
AI Technical Summary
Existing technologies struggle to extract features from the massive amounts of telemetry data from spacecraft that can effectively characterize the performance degradation of critical equipment, resulting in insufficient accuracy in health management research.
A multi-domain physical feature extraction method was adopted, combined with stacked autoencoders and correlation analysis. Performance characterization and evaluation indicators were constructed using Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient. Feature sets that can describe equipment performance degradation were screened from spacecraft telemetry data.
It effectively extracted features that characterize the performance degradation of key spacecraft equipment, improving the accuracy and reliability of health management research.
Smart Images
Figure CN121858952A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for screening high-dimensional spacecraft data features based on multivariate feature extraction, belonging to the aerospace field. Background Technology
[0002] With the rapid development of aerospace technology, conducting health management research is crucial to ensuring the safe, reliable, and stable operation of spacecraft in orbit. The effectiveness of features extracted from raw telemetry data directly determines the accuracy of subsequent health management studies, such as fault diagnosis or health assessment. Therefore, extracting state features that effectively characterize the performance degradation of critical equipment from complex, high-dimensional telemetry data is of significant research importance. However, critical equipment on spacecraft often has multiple monitoring sites, each generating massive amounts of telemetry data. Existing feature extraction methods struggle to extract features that effectively characterize the performance evolution of critical equipment from this vast amount of telemetry data. Summary of the Invention
[0003] The technical problem solved by this invention is to overcome the shortcomings of existing technologies and provide a method for high-dimensional data feature screening of spacecraft based on multi-dimensional feature extraction. This method, through multi-domain physical feature extraction, stacked autoencoders, and correlation analysis, obtains a set of features that can effectively describe the performance degradation of key spacecraft equipment, thus solving the problem that existing feature extraction methods struggle to extract effective degradation features from massive amounts of telemetry data.
[0004] The technical solution of this invention is:
[0005] A method for feature filtering of high-dimensional spacecraft data based on multivariate feature extraction, comprising the following steps:
[0006] (1) Collect the original telemetry sequences of m telemetry sites of key equipment of spacecraft. The original telemetry sequence of each telemetry site is composed of the telemetry data collected at time point 1, time point 2, ..., time point n, arranged in order of the collection time from earliest to latest. Preprocess the original telemetry sequences of all telemetry sites and construct a telemetry dataset using the telemetry data of all telemetry sites after preprocessing.
[0007] (2) Based on the multi-domain physical feature formula, obtain the set of explicit physical features for each telemetry point from the telemetry dataset; the set of physical features includes the maximum value sequence, minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root mean square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, margin factor sequence, centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence;
[0008] (3) Using a stacked autoencoder, the latent association feature sequence of multidimensional telemetry sites is obtained from the telemetry dataset;
[0009] (4) Using Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient, construct performance characterization evaluation index; use performance characterization evaluation index to screen the feature sequences in the set of explicit physical features of each telemetry site, merge the implicit correlation feature sequences and the screened set of explicit physical features together, and then construct the performance degradation feature set of the multidimensional telemetry sites of the key equipment of the spacecraft.
[0010] Furthermore, the step of preprocessing the original telemetry sequences of all telemetry sites in step (1) is as follows:
[0011] (1.1) Interpolate and compensate the telemetry sequence of each telemetry site to obtain the interpolated compensated telemetry sequence corresponding to each telemetry site;
[0012] (1.2) Based on the Wright criterion, outlier removal is performed on the interpolated compensated telemetry sequence to obtain the final telemetry sequence for each telemetry site, thus completing the preprocessing step.
[0013] Furthermore, the process of extracting explicit physical features from the telemetry dataset based on the multi-domain physical feature formula in step (2) is as follows:
[0014] (2.1) Using the preprocessed telemetry dataset, obtain the time-domain feature set for each telemetry point; the time-domain feature set includes the maximum value sequence, minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, and margin factor sequence;
[0015] (2.2) Using the preprocessed telemetry dataset, obtain the frequency domain feature set for each telemetry point; the frequency domain feature set includes the centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence.
[0016] (2.3) The time-domain feature set and frequency-domain feature set corresponding to each telemetry site are merged together to form the explicit physical feature set of that telemetry site.
[0017] Furthermore, the specific process for obtaining the temporal feature set of each telemetry point in step (2.1) is as follows:
[0018] (2.1.1) Perform sliding window processing on the preprocessed telemetry sequence for each telemetry site; the sliding window processing is to slide along the order of the telemetry sequence with a sliding window of size L, and form a sliding window sequence with every L consecutive values; if the number of elements at the end of the telemetry sequence is less than L, then the elements at the end of the telemetry sequence are divided into a separate sliding window sequence, and the data volume is the number of elements at the end; all sliding window sequences form the sliding window sequence set for the corresponding telemetry site.
[0019] (2.1.2) Calculate the maximum, minimum, average, peak-to-peak value, rectified average, variance, standard deviation, root mean square (RMS), root square amplitude, kurtosis, skewness, waveform factor, peak-to-peak value factor, impulse factor, and margin factor for each sliding window sequence in the sliding window sequence set corresponding to each telemetry point; construct the maximum and minimum value sequences for that telemetry point by combining the maximum, minimum, average, peak-to-peak value, rectified average, variance, standard deviation, RMS, root square amplitude, kurtosis, skewness, waveform factor, peak-to-peak value factor, impulse factor, and margin factor of all sliding window sequences in the sliding window sequence set of each telemetry point. The sequence includes: average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, and margin factor sequence; the maximum value sequence, minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, and margin factor sequence of each telemetry point constitute the time-domain feature set of that telemetry point.
[0020] The formula for calculating the average value of the sliding window sequence is as follows: The formula for calculating the rectified average value is: The formula for calculating variance is: The formula for calculating the root mean square is: The formula for calculating the root amplitude is as follows: The formula for calculating kurtosis is: The formula for calculating skewness is: The formula for calculating the waveform factor is: The formula for calculating the peak-to-peak factor is: The formula for calculating the impulse factor is: Where, x i For the i-th element in the sliding window sequence, max(x) i ) represents the maximum value in the sliding window sequence.
[0021] Furthermore, the specific process for obtaining the frequency domain feature set of each telemetry point in step (2.2) is as follows:
[0022] (2.2.1) Perform sliding window processing on the preprocessed telemetry sequence for each telemetry site; the sliding window processing is to slide along the order of the telemetry sequence with a sliding window of size L, and form a sliding window sequence with every L consecutive values; if the number of elements at the end of the telemetry sequence is less than L, then the several elements at the end of the telemetry sequence are divided into a separate sliding window sequence, and the data volume is the number of elements at the end; all sliding window sequences form the sliding window sequence set for the corresponding telemetry site.
[0023] (2.2.2) Perform a Fourier transform on each sliding window sequence in the set of sliding window sequences corresponding to each telemetry point to obtain the spectrum corresponding to each sliding window sequence;
[0024] (2.2.3) Based on the spectrum corresponding to each sliding window sequence, calculate the centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation of each sliding window sequence; combine the centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation of all sliding window sequences in the set of sliding window sequences of each telemetry point to form the centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence of that telemetry point; the centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence of each telemetry point form the frequency domain feature set of that telemetry point;
[0025] The formula for calculating the centroid frequency of the sliding window sequence is as follows: The formula for calculating mean square frequency is: The formula for calculating the root mean square frequency is: The formula for calculating frequency variance is: The formula for calculating the standard deviation of frequency is: Where y(k) is the spectrum, f k Let be the frequency value of the k-th spectral line.
[0026] Furthermore, the specific steps for obtaining the latent association feature sequence of multidimensional telemetry sites from the telemetry dataset in step (3) are as follows:
[0027] (3.1) Input the telemetry sequences of all telemetry sites into a stacked autoencoder; the stacked autoencoder has three layers of encoders and three layers of decoders; the stacked autoencoder associates and couples the telemetry data collected at the same time point of the first telemetry site, the second telemetry site, ..., the mth telemetry site, and outputs the initial implicit association features at the corresponding collection time; the initial implicit association features at all times are combined together to form a multidimensional sequence of initial implicit association features of telemetry sites with a length of n;
[0028] (3.2) Perform sliding window processing on the initial implicit association feature sequence of the multidimensional telemetry site; the sliding window processing is to slide along the order of the initial implicit association feature sequence with a sliding window of size L, and form a hidden sliding window sequence with every L consecutive values; if the number of end elements of the initial implicit association feature sequence is less than L, then the several elements at the end of the initial implicit association feature sequence are divided into a separate sliding window sequence, and the data volume is the number of end elements; all the hidden sliding window sequences form the set of hidden sliding window sequences of the multidimensional telemetry site.
[0029] (3.3) Calculate the average value of each hidden sliding window sequence in the set of hidden sliding window sequences of multidimensional telemetry sites, and form the hidden association feature sequence of multidimensional telemetry sites by averaging the values of all hidden sliding window sequences.
[0030] The specific steps for filtering the feature sequences in the set of explicit physical features of each telemetry site using performance characterization evaluation indicators are as follows:
[0031] (4.1) Calculate the Pearson correlation coefficient between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent association feature sequence of the multidimensional telemetry site;
[0032] (4.2) Calculate the Spearman correlation coefficient between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent association feature sequence of the multidimensional telemetry site;
[0033] (4.3) Calculate the Kendall correlation coefficient between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent association feature sequence of the multidimensional telemetry site;
[0034] (4.4) Calculate the performance characterization evaluation index based on the Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient between the maximum value sequence in the first telemetry site's explicit physical feature set and the implicit correlation feature sequence of the multidimensional telemetry site;
[0035] (4.5) Screen the maximum value sequence in the set of explicit physical characteristics of the first telemetry site. If the performance characterization evaluation index is greater than 0.5, retain the maximum value sequence; otherwise, remove the maximum value sequence.
[0036] (4.6) Perform feature screening on the minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root mean square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, margin factor sequence, centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence in the set of explicit physical characteristics of the first telemetry site. The specific process is the same as steps (4.1) to (4.5).
[0037] (4.7) Perform feature screening on the feature sequences in the set of dominant physical features of the second telemetry site, the third telemetry site, ..., the mth telemetry site. The specific process is the same as steps (4.1) to (4.6).
[0038] Furthermore, in step (4.1), the Pearson correlation coefficient ρ between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent association feature sequence of the multidimensional telemetry site... p The calculation formula is:
[0039]
[0040] Among them, F i Let Y be the sequence of maximum values, and let Y be the sequence of latent association features. σ is the variance of the sequence of maximum values. Y Let cov(F) be the variance of the latent association feature sequences. i Y) represents the covariance between the maximum value sequence and the latent association feature sequence;
[0041] The calculation process for the Spearman correlation coefficient between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent correlation feature sequence of the multidimensional telemetry site in step (4.2) is as follows:
[0042] First, the elements in the maximum value sequence and the latent association feature sequence are sorted separately, so that the elements in the maximum value sequence and the latent association feature sequence are arranged from smallest to largest, thus obtaining the ordered maximum value sequence and the ordered latent association feature sequence.
[0043] Next, a level is assigned to each element in the ordered maximum value sequence and the ordered implicit association feature sequence. The rule for assigning levels to elements is that the index value of the element in the sequence is used as the level of the element, that is, the level of the smallest element in the ordered maximum value sequence and the ordered implicit association feature sequence is 1, the level of the largest element is N, and N is the sequence length.
[0044] Finally, the Spearman correlation coefficient ρ was calculated. s The formula is
[0045]
[0046] Where, r i The difference in rank between the i-th element in the maximum value sequence and the latent association feature sequence;
[0047] The Kendall correlation coefficient between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent correlation feature sequence of the multidimensional telemetry site in step (4.3) is calculated as follows:
[0048] First, all elements in the maximum value sequence and the latent association feature sequence are paired into N element pairs, where the N element pairs are (a1, b1), (a2, b2), ..., (a... N ,b N );
[0049] Then, select all pairs of elements that satisfy 1≤i<j≤N to form [(a i ,b i ), (a j ,b j ]], forming (N(N-1)) / 2 unordered data pairs; where a i and b i These are the i-th elements in the maximum value sequence and the latent association feature sequence, respectively, a j and b j These are the j-th elements in the maximum value sequence and the latent association feature sequence, respectively; perform a consistency check on all unordered data pairs, if (a i -a j (b) i -b j If the value is greater than 0, the group is a consistent pair; otherwise, the group is a inconsistent pair. The total number of consistent pairs and inconsistent pairs are accumulated and represented as H and D, respectively.
[0050] Finally, the Kendall correlation coefficient τ is calculated using the following formula:
[0051]
[0052] Furthermore, the calculation formula for the performance characterization evaluation index P in step (4.4) is as follows:
[0053] P = |0.4 × ρ p +0.3×ρ s +0.3×τ|
[0054] Where, ρ p ρ s τ and τ represent the Pearson correlation coefficient, Spearman correlation coefficient, and Kendall correlation coefficient, respectively.
[0055] The advantages of this invention compared to the prior art are:
[0056] (1) This invention obtains a set of features from massive telemetry data through multi-domain physical feature extraction, stacked autoencoders and correlation analysis. The features in this set can effectively characterize the performance degradation of key spacecraft equipment.
[0057] (2) This invention extracts features from telemetry data from multiple perspectives, such as time domain signal and frequency domain energy, through multi-domain physical feature formulas, and further utilizes stacked autoencoders to mine the implicit relationships between coupled telemetry data, thereby obtaining sufficient and comprehensive features that can describe the performance evolution process of key spacecraft equipment.
[0058] (3) This invention uses three correlation coefficients, namely Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient, to construct a performance characterization evaluation index formula. It quantifies the effectiveness of the extracted features from the perspective of three correlation analyses, thereby constructing a feature set that effectively describes the performance degradation of key spacecraft equipment. Attached Figure Description
[0059] Figure 1 This is a flowchart of a spacecraft high-dimensional data feature screening method based on multivariate feature extraction according to the present invention;
[0060] Figure 2 This is a diagram of the stacked autoencoder network structure used in the spacecraft high-dimensional data feature screening method based on multivariate feature extraction of the present invention. Detailed Implementation
[0061] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.
[0062] like Figure 1 As shown, this invention provides a method for feature filtering of high-dimensional spacecraft data based on multivariate feature extraction, the steps of which are as follows:
[0063] (1) Collect the original telemetry sequences of m telemetry sites of key equipment of spacecraft. The original telemetry sequence of each telemetry site is composed of the telemetry data collected at time point 1, time point 2, ..., time point n, arranged in order of the collection time from earliest to latest. Preprocess the original telemetry sequences of all telemetry sites and construct a telemetry dataset using the telemetry data of all telemetry sites after preprocessing.
[0064] The step of preprocessing the original telemetry sequences of all telemetry sites in step (1) is as follows:
[0065] (1.1) Interpolate and compensate the telemetry sequence of each telemetry site to obtain the interpolated compensated telemetry sequence corresponding to each telemetry site;
[0066] (1.2) Based on the Wright criterion, outlier removal is performed on the interpolated compensated telemetry sequence to obtain the final telemetry sequence for each telemetry site, thus completing the preprocessing step.
[0067] (2) Based on the multi-domain physical feature formula, obtain the set of explicit physical features for each telemetry point from the telemetry dataset; the set of physical features includes the maximum value sequence, minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root mean square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, margin factor sequence, centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence;
[0068] The process of extracting explicit physical features from the telemetry dataset based on the multi-domain physical feature formula in step (2) is as follows:
[0069] (2.1) Using the preprocessed telemetry dataset, obtain the time-domain feature set for each telemetry point; the time-domain feature set includes the maximum value sequence, minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, and margin factor sequence;
[0070] The specific process for obtaining the temporal feature set of each telemetry point in step (2.1) is as follows:
[0071] (2.1.1) Perform sliding window processing on the preprocessed telemetry sequence for each telemetry site; the sliding window processing is to slide along the order of the telemetry sequence with a sliding window of size L, and form a sliding window sequence with every L consecutive values; if the number of elements at the end of the telemetry sequence is less than L, then the elements at the end of the telemetry sequence are divided into a separate sliding window sequence, and the data volume is the number of elements at the end; all sliding window sequences form the sliding window sequence set for the corresponding telemetry site.
[0072] (2.1.2) Calculate the maximum, minimum, average, peak-to-peak value, rectified average, variance, standard deviation, root mean square (RMS), root square amplitude, kurtosis, skewness, waveform factor, peak-to-peak value factor, impulse factor, and margin factor for each sliding window sequence in the sliding window sequence set corresponding to each telemetry point; construct the maximum and minimum value sequences for that telemetry point by combining the maximum, minimum, average, peak-to-peak value, rectified average, variance, standard deviation, RMS, root square amplitude, kurtosis, skewness, waveform factor, peak-to-peak value factor, impulse factor, and margin factor of all sliding window sequences in the sliding window sequence set of each telemetry point. The sequence includes: average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, and margin factor sequence; the maximum value sequence, minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, and margin factor sequence of each telemetry point constitute the time-domain feature set of that telemetry point.
[0073] The formula for calculating the average value of the sliding window sequence is as follows: The formula for calculating the rectified average value is: The formula for calculating variance is: The formula for calculating the root mean square is: The formula for calculating the root amplitude is as follows: The formula for calculating kurtosis is: The formula for calculating skewness is: The formula for calculating the waveform factor is: The formula for calculating the peak-to-peak factor is: The formula for calculating the impulse factor is: Where, x i For the i-th element in the sliding window sequence, max(x) i ) represents the maximum value in the sliding window sequence.
[0074] (2.2) Using the preprocessed telemetry dataset, obtain the frequency domain feature set for each telemetry point; the frequency domain feature set includes the centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence.
[0075] The specific process for obtaining the frequency domain feature set of each telemetry point in step (2.2) is as follows:
[0076] (2.2.1) Perform sliding window processing on the preprocessed telemetry sequence for each telemetry site; the sliding window processing is to slide along the order of the telemetry sequence with a sliding window of size L, and form a sliding window sequence with every L consecutive values; if the number of elements at the end of the telemetry sequence is less than L, then the several elements at the end of the telemetry sequence are divided into a separate sliding window sequence, and the data volume is the number of elements at the end; all sliding window sequences form the sliding window sequence set for the corresponding telemetry site.
[0077] (2.2.2) Perform a Fourier transform on each sliding window sequence in the set of sliding window sequences corresponding to each telemetry point to obtain the spectrum corresponding to each sliding window sequence;
[0078] (2.2.3) Based on the spectrum corresponding to each sliding window sequence, calculate the centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation of each sliding window sequence; combine the centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation of all sliding window sequences in the set of sliding window sequences of each telemetry point to form the centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence of that telemetry point; the centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence of each telemetry point form the frequency domain feature set of that telemetry point;
[0079] The formula for calculating the centroid frequency of the sliding window sequence is as follows: The formula for calculating mean square frequency is: The formula for calculating the root mean square frequency is: The formula for calculating frequency variance is: The formula for calculating the standard deviation of frequency is: Where y(k) is the spectrum, f k Let be the frequency value of the k-th spectral line.
[0080] (2.3) The time-domain feature set and frequency-domain feature set corresponding to each telemetry site are merged together to form the explicit physical feature set of that telemetry site.
[0081] (3) Using a stacked autoencoder, the latent association feature sequence of multidimensional telemetry sites is obtained from the telemetry dataset. The specific process is as follows:
[0082] (3.1) Input the telemetry sequences of all telemetry points into the stacked autoencoder; such as Figure 2As shown, the stacked autoencoder has three layers of encoders and three layers of decoders. The network structure configuration is shown in Table 1, where INPUT is the dimension of the telemetry dataset. The stacked autoencoder associates and couples the telemetry data collected at the same time point from the first telemetry point, the second telemetry point, ..., the m-th telemetry point, and outputs the initial implicit association features at the corresponding collection time. The initial implicit association features at all times are combined to form a multidimensional sequence of initial implicit association features of the telemetry point with a length of n.
[0083] Table 1 Stacked Autoencoder Network Configuration
[0084] number of floors Input data dimensionality Output data dimensionality Activation function Encoder layer1 INPUT 64 Relu Encoder layer2 64 32 Relu Encoder layer3 32 1 Relu Decoder layer1 1 32 Relu Decoder layer2 32 64 Relu Decoder layer3 64 INPUT Sigmoid
[0085] (3.2) Perform sliding window processing on the initial implicit association feature sequence of the multidimensional telemetry site; the sliding window processing is to slide along the order of the initial implicit association feature sequence with a sliding window of size L, and form a hidden sliding window sequence with every L consecutive values; if the number of end elements of the initial implicit association feature sequence is less than L, then the several elements at the end of the initial implicit association feature sequence are divided into a separate sliding window sequence, and the data volume is the number of end elements; all the hidden sliding window sequences form the set of hidden sliding window sequences of the multidimensional telemetry site.
[0086] (3.3) Calculate the average value of each hidden sliding window sequence in the set of hidden sliding window sequences of multidimensional telemetry sites, and form the hidden association feature sequence of multidimensional telemetry sites by averaging the values of all hidden sliding window sequences.
[0087] Through the above steps, this invention extracts features from telemetry data from multiple perspectives, such as time-domain signals and frequency-domain energy, based on multi-domain physical feature formulas. Furthermore, it utilizes stacked autoencoders to mine the implicit relationships between coupled telemetry data, thereby obtaining sufficient and comprehensive features that can describe the performance evolution process of key spacecraft equipment.
[0088] (4) Using Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient, construct performance characterization evaluation index; use performance characterization evaluation index to screen the feature sequences in the set of explicit physical features of each telemetry site, merge the implicit correlation feature sequences and the screened set of explicit physical features together, and then construct the performance degradation feature set of the multidimensional telemetry sites of the key equipment of the spacecraft.
[0089] The specific steps for filtering the feature sequences in the set of explicit physical features of each telemetry site using performance characterization evaluation indicators are as follows:
[0090] (4.1) Calculate the Pearson correlation coefficient ρ between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent association feature sequence of the multidimensional telemetry site.p The calculation formula is:
[0091]
[0092] Among them, F i Let Y be the sequence of maximum values, and let Y be the sequence of latent association features. σ is the variance of the sequence of maximum values. Y Let cov(F) be the variance of the latent association feature sequences. i Y) represents the covariance between the maximum value sequence and the latent association feature sequence;
[0093] (4.2) Calculate the Spearman correlation coefficient ρ between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent correlation feature sequence of the multidimensional telemetry site. s The calculation process is as follows:
[0094] First, the elements in the maximum value sequence and the latent association feature sequence are sorted separately, so that the elements in the maximum value sequence and the latent association feature sequence are arranged from smallest to largest, thus obtaining the ordered maximum value sequence and the ordered latent association feature sequence.
[0095] Next, a level is assigned to each element in the ordered maximum value sequence and the ordered implicit association feature sequence. The rule for assigning levels to elements is that the index value of the element in the sequence is used as the level of the element, that is, the level of the smallest element in the ordered maximum value sequence and the ordered implicit association feature sequence is 1, the level of the largest element is N, and N is the sequence length.
[0096] Finally, the Spearman correlation coefficient ρ was calculated. s The formula is
[0097]
[0098] Where, r i The difference in rank between the i-th element in the maximum value sequence and the latent association feature sequence;
[0099] (4.3) Calculate the Kendall correlation coefficient τ between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent association feature sequence of the multidimensional telemetry site. The calculation process is as follows:
[0100] First, all elements in the maximum value sequence and the latent association feature sequence are paired into N element pairs, where the N element pairs are (a1, b1), (a2, b2), ..., (a... N ,b N );
[0101] Then, select all pairs of elements that satisfy 1≤i<j≤N to form [(a i,b i ), (a j ,b j ]], forming (N(N-1)) / 2 unordered data pairs; where a i and b i These are the i-th elements in the maximum value sequence and the latent association feature sequence, respectively, a j and b j These are the j-th elements in the maximum value sequence and the latent association feature sequence, respectively; perform a consistency check on all unordered data pairs, if (a i -a j (b) i -b j If the value is greater than 0, the group is a consistent pair; otherwise, the group is a inconsistent pair. The total number of consistent pairs and inconsistent pairs are accumulated and represented as H and D, respectively.
[0102] Finally, the Kendall correlation coefficient τ is calculated using the following formula:
[0103]
[0104] (4.4) Based on the Pearson correlation coefficient, Spearman correlation coefficient, and Kendall correlation coefficient between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent correlation feature sequence of the multidimensional telemetry site, the performance characterization evaluation index is calculated. The calculation formula is as follows:
[0105] P = |0.4 × ρ p +0.3×ρ s +0.3×τ|.
[0106] (4.5) Screen the maximum value sequence in the set of explicit physical characteristics of the first telemetry site. If the performance characterization evaluation index is greater than 0.5, retain the maximum value sequence; otherwise, remove the maximum value sequence.
[0107] (4.6) Perform feature screening on the minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root mean square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, margin factor sequence, centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence in the set of explicit physical characteristics of the first telemetry site. The specific process is the same as steps (4.1) to (4.5).
[0108] (4.7) Feature screening is performed on the feature sequences in the set of explicit physical features of the second telemetry site, the third telemetry site, ..., the mth telemetry site. The specific process is the same as steps (4.1) to (4.6). Through the above steps, this invention uses three correlation coefficients, namely Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient, to construct a performance characterization evaluation index formula. The effectiveness of the extracted features is quantified from the perspective of three correlation analyses, thereby constructing a feature set that effectively describes the performance degradation of key spacecraft equipment.
[0109] In summary, this invention obtains a set of features from massive telemetry data through multi-domain physical feature extraction, stacked autoencoders, and correlation analysis. The features in this set can effectively characterize the performance degradation of key spacecraft equipment.
[0110] The parts of this invention not described in detail are common knowledge to those skilled in the art.
Claims
1. A method for feature selection of high-dimensional spacecraft data based on multivariate feature extraction, characterized in that... The steps are as follows: (1) Collect the original telemetry sequences of m telemetry sites of key equipment of spacecraft. The original telemetry sequence of each telemetry site is composed of the telemetry data collected at time point 1, time point 2, ..., time point n, arranged in order of the collection time from earliest to latest. Preprocess the original telemetry sequences of all telemetry sites and construct a telemetry dataset using the telemetry data of all telemetry sites after preprocessing. (2) Based on the multi-domain physical feature formula, obtain the set of explicit physical features for each telemetry point from the telemetry dataset; the set of physical features includes the maximum value sequence, minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root mean square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, margin factor sequence, centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence; (3) Using a stacked autoencoder, the latent association feature sequence of multidimensional telemetry sites is obtained from the telemetry dataset; (4) Using Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient, construct performance characterization evaluation index; use performance characterization evaluation index to screen the feature sequences in the set of explicit physical features of each telemetry site, merge the implicit correlation feature sequences and the screened set of explicit physical features together, and then construct the performance degradation feature set of the multidimensional telemetry sites of the key equipment of the spacecraft.
2. The method for screening high-dimensional spacecraft data features based on multivariate feature extraction according to claim 1, characterized in that: The step of preprocessing the original telemetry sequences of all telemetry sites in step (1) is as follows: (1.1) Interpolate and compensate the telemetry sequence of each telemetry site to obtain the interpolated compensated telemetry sequence corresponding to each telemetry site; (1.2) Based on the Wright criterion, outlier removal is performed on the interpolated compensated telemetry sequence to obtain the final telemetry sequence for each telemetry site, thus completing the preprocessing step.
3. The spacecraft high-dimensional data feature filtering method based on multivariate feature extraction according to claim 1, characterized in that: The process of extracting explicit physical features from the telemetry dataset based on the multi-domain physical feature formula in step (2) is as follows: (2.1) Using the preprocessed telemetry dataset, obtain the time-domain feature set for each telemetry point; the time-domain feature set includes the maximum value sequence, minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, and margin factor sequence; (2.2) Using the preprocessed telemetry dataset, obtain the frequency domain feature set for each telemetry point; the frequency domain feature set includes the centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence. (2.3) The time-domain feature set and frequency-domain feature set corresponding to each telemetry site are merged together to form the explicit physical feature set of that telemetry site.
4. The method for screening high-dimensional spacecraft data features based on multivariate feature extraction according to claim 3, characterized in that: The specific process for obtaining the temporal feature set of each telemetry point in step (2.1) is as follows: (2.1.1) Perform sliding window processing on the preprocessed telemetry sequence for each telemetry site; the sliding window processing is to slide along the order of the telemetry sequence with a sliding window of size L, and form a sliding window sequence with every L consecutive values; if the number of elements at the end of the telemetry sequence is less than L, then the elements at the end of the telemetry sequence are divided into a separate sliding window sequence, and the data volume is the number of elements at the end; all sliding window sequences form the sliding window sequence set for the corresponding telemetry site. (2.1.2) Calculate the maximum, minimum, average, peak-to-peak value, rectified average, variance, standard deviation, root mean square (RMS), root square amplitude, kurtosis, skewness, waveform factor, peak-to-peak value factor, impulse factor, and margin factor for each sliding window sequence in the sliding window sequence set corresponding to each telemetry point; construct the maximum and minimum value sequences for that telemetry point by combining the maximum, minimum, average, peak-to-peak value, rectified average, variance, standard deviation, RMS, root square amplitude, kurtosis, skewness, waveform factor, peak-to-peak value factor, impulse factor, and margin factor of all sliding window sequences in the sliding window sequence set of each telemetry point. The sequence includes: average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, and margin factor sequence; the maximum value sequence, minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, and margin factor sequence of each telemetry point constitute the time-domain feature set of that telemetry point. The formula for calculating the average value of the sliding window sequence is as follows: The formula for calculating the rectified average value is: The formula for calculating variance is: The formula for calculating the root mean square is: The formula for calculating the root amplitude is as follows: The formula for calculating kurtosis is: The formula for calculating skewness is: The formula for calculating the waveform factor is: The formula for calculating the peak-to-peak factor is: The formula for calculating the impulse factor is: Where, x i For the i-th element in the sliding window sequence, max(x) i ) represents the maximum value in the sliding window sequence.
5. The spacecraft high-dimensional data feature screening method based on multivariate feature extraction according to claim 3, characterized in that: The specific process for obtaining the frequency domain feature set of each telemetry point in step (2.2) is as follows: (2.2.1) Perform sliding window processing on the preprocessed telemetry sequence for each telemetry site; the sliding window processing is to slide along the order of the telemetry sequence with a sliding window of size L, and form a sliding window sequence with every L consecutive values; if the number of elements at the end of the telemetry sequence is less than L, then the several elements at the end of the telemetry sequence are divided into a separate sliding window sequence, and the data volume is the number of elements at the end; all sliding window sequences form the sliding window sequence set for the corresponding telemetry site. (2.2.2) Perform a Fourier transform on each sliding window sequence in the set of sliding window sequences corresponding to each telemetry point to obtain the spectrum corresponding to each sliding window sequence; (2.2.3) Based on the spectrum corresponding to each sliding window sequence, calculate the centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation of each sliding window sequence; combine the centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation of all sliding window sequences in the set of sliding window sequences of each telemetry point to form the centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence of that telemetry point; the centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence of each telemetry point form the frequency domain feature set of that telemetry point; The formula for calculating the centroid frequency of the sliding window sequence is as follows: The formula for calculating mean square frequency is: The formula for calculating the root mean square frequency is: The formula for calculating frequency variance is: The formula for calculating the standard deviation of frequency is: Where y(k) is the spectrum, f k Let be the frequency value of the k-th spectral line.
6. The method for screening high-dimensional spacecraft data features based on multivariate feature extraction according to claim 1, characterized in that: The specific steps for obtaining the latent association feature sequence of multidimensional telemetry sites from the telemetry dataset in step (3) are as follows: (3.1) Input the telemetry sequences of all telemetry sites into the stacked autoencoder; the stacked autoencoder is equipped with three layers of encoders and three layers of decoders; the stacked autoencoder associates and couples the telemetry data collected at the same time point of the first telemetry site, the second telemetry site, ..., the mth telemetry site, and outputs the initial implicit association features at the corresponding collection time. The initial latent association features at all times are combined to form a sequence of initial latent association features of a multidimensional telemetry site of length n. (3.2) Perform sliding window processing on the initial implicit association feature sequence of the multidimensional telemetry site; the sliding window processing is to slide along the order of the initial implicit association feature sequence with a sliding window of size L, and form a hidden sliding window sequence with every L consecutive values; if the number of end elements of the initial implicit association feature sequence is less than L, then the several elements at the end of the initial implicit association feature sequence are divided into a separate sliding window sequence, and the data volume is the number of end elements; all the hidden sliding window sequences form the set of hidden sliding window sequences of the multidimensional telemetry site. (3.3) Calculate the average value of each hidden sliding window sequence in the set of hidden sliding window sequences of multidimensional telemetry sites, and form the hidden association feature sequence of multidimensional telemetry sites by averaging the values of all hidden sliding window sequences.
7. The method for screening high-dimensional spacecraft data features based on multivariate feature extraction according to claim 1, characterized in that: The specific steps for filtering the feature sequences in the set of explicit physical features of each telemetry site using performance characterization evaluation indicators are as follows: (4.1) Calculate the Pearson correlation coefficient between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent association feature sequence of the multidimensional telemetry site; (4.2) Calculate the Spearman correlation coefficient between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent association feature sequence of the multidimensional telemetry site; (4.3) Calculate the Kendall correlation coefficient between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent association feature sequence of the multidimensional telemetry site; (4.4) Calculate the performance characterization evaluation index based on the Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient between the maximum value sequence in the first telemetry site's explicit physical feature set and the implicit correlation feature sequence of the multidimensional telemetry site; (4.5) Screen the maximum value sequence in the set of explicit physical characteristics of the first telemetry site. If the performance characterization evaluation index is greater than 0.5, retain the maximum value sequence; otherwise, remove the maximum value sequence. (4.6) Perform feature screening on the minimum value sequence, average value sequence, peak-to-peak value sequence, rectified average value sequence, variance sequence, standard deviation sequence, root mean square sequence, root mean square amplitude sequence, kurtosis sequence, skewness sequence, waveform factor sequence, peak-to-peak value factor sequence, impulse factor sequence, margin factor sequence, centroid frequency sequence, mean square frequency sequence, root mean square frequency sequence, frequency variance sequence, and frequency standard deviation sequence in the set of explicit physical characteristics of the first telemetry site. The specific process is the same as steps (4.1) to (4.5). (4.7) Perform feature screening on the feature sequences in the set of dominant physical features of the second telemetry site, the third telemetry site, ..., the mth telemetry site. The specific process is the same as steps (4.1) to (4.6).
8. The method for screening high-dimensional spacecraft data features based on multivariate feature extraction according to claim 7, characterized in that: In step (4.1), the Pearson correlation coefficient ρ between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent association feature sequence of the multidimensional telemetry site p The calculation formula is: Among them, F i Y is the sequence of maximum values, and Y is the sequence of latent association features. σ is the variance of the sequence of maximum values. Y Let cov(F) be the variance of the latent association feature sequences. i Y) represents the covariance between the maximum value sequence and the latent association feature sequence; The calculation process for the Spearman correlation coefficient between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent correlation feature sequence of the multidimensional telemetry site in step (4.2) is as follows: First, the elements in the maximum value sequence and the latent association feature sequence are sorted separately, so that the elements in the maximum value sequence and the latent association feature sequence are arranged from smallest to largest, thus obtaining the ordered maximum value sequence and the ordered latent association feature sequence. Next, a level is assigned to each element in the ordered maximum value sequence and the ordered implicit association feature sequence. The rule for assigning levels to elements is that the index value of the element in the sequence is used as the level of the element, that is, the level of the smallest element in the ordered maximum value sequence and the ordered implicit association feature sequence is 1, the level of the largest element is N, and N is the sequence length. Finally, the Spearman correlation coefficient ρ was calculated. s The formula is Where, r i The difference in rank between the i-th element in the maximum value sequence and the latent association feature sequence; The Kendall correlation coefficient between the maximum value sequence in the dominant physical feature set of the first telemetry site and the latent correlation feature sequence of the multidimensional telemetry site in step (4.3) is calculated as follows: First, all elements in the maximum value sequence and the latent association feature sequence are paired into N element pairs, where the N element pairs are (a1, b1), (a2, b2), ..., (a... N ,b N ); Then, select all pairs of elements that satisfy 1≤i<j≤N to form [(a i ,b i ), (a j ,b j ]], forming (N(N-1)) / 2 unordered data pairs; where a i and b i These are the i-th elements in the maximum value sequence and the latent association feature sequence, respectively, a j and b j These are the j-th elements in the maximum value sequence and the latent association feature sequence, respectively; perform a consistency check on all unordered data pairs, if (a i -a j (b) i -b j If the value is greater than 0, the group is a consistent pair; otherwise, the group is a inconsistent pair. The total number of consistent pairs and inconsistent pairs are accumulated and represented as H and D, respectively. Finally, the Kendall correlation coefficient τ is calculated using the following formula:
9. A method for screening high-dimensional spacecraft data features based on multivariate feature extraction according to claim 7, characterized in that: The formula for calculating the performance characterization evaluation index P in step (4.4) is as follows: P=|0.4×ρ p +0.3×ρ s +0.3×τ| Where, ρ p ρs and τ are the Pearson correlation coefficient, Spearman correlation coefficient, and Kendall correlation coefficient, respectively.