Center-constrained contrastive learning feature transformation self-integrated satellite anomaly detection method

Abstract

The application discloses a center-constrained contrast learning feature transformation self-integrated satellite anomaly detection method, and the method comprises the following steps: receiving telemetry original data transmitted by a spacecraft; extracting the telemetry original data according to a spacecraft load single-machine device to be judged to obtain corresponding subpackage telemetry data; inputting the preprocessed subpackage telemetry data into a pre-established and pre-trained anomaly discrimination model to realize load single-machine device state discrimination; the anomaly discrimination model is based on a center-constrained and contrast-constrained learning self-integrated anomaly detection method, feature mapping is realized by fusing a center loss and a contrast loss, and the self-integrated learning mode of multi-view and multi-level features is adopted to realize spacecraft telemetry data anomaly detection. The application can be universally applied to load single-machine device state discrimination learning and automatic state recognition for spacecraft task targets, and the adaptability and automation level are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the application fields of satellite and payload on-orbit test and operation management and control, instrument monitoring, etc. In particular, it relates to a self-integrated satellite anomaly detection method based on center-constrained contrast learning feature transformation. BACKGROUND

[0002] During the on-orbit operation of a satellite, timely and effective detection and identification of anomalies and taking effective measures for maintenance are of great significance to ensure the safe, reliable and efficient operation of the satellite. Telemetry parameter data is the only basis for experts and scholars and ground operation and management personnel of a spacecraft to understand the on-orbit state of the spacecraft instruments during on-orbit test and on-orbit operation. Data-driven anomaly detection methods use statistical analysis, data mining and machine learning methods to model telemetry data, and use the obtained model to find anomalies, which has good scalability and adaptability.

[0003] With the increase of the national comprehensive strength and the improvement of the scientific and technological level, on the one hand, spacecraft and payload equipment and instruments are more personalized, precise and complex with tasks, and single-view and single-layer features can lead to loss of abnormal decision information, cannot reflect the true state of the satellite, and are easily disturbed by noise data, resulting in poor anomaly detection effect; on the other hand, with the increasing number of on-orbit tasks, the function design is complex, the format of the downlink telemetry parameters is diverse, and the dimension of the telemetry parameters reaches thousands, the irregular changes of the telemetry parameters caused by anomalies, and the distinction between normal and abnormal samples is not obvious. These bring great difficulty to the spacecraft equipment state anomaly discrimination of limited arc tracking resources and human resources, and pose a challenge to the anomaly detection algorithm.

[0004] Traditional parameter interpretation methods rely on expert knowledge and cannot meet the increasingly complex task requirements, and have low interpretation efficiency and data utilization rate, and also have problems such as misjudgment and omission. At present, domestic and foreign experts and scholars have made in-depth research on telemetry parameters through spacecraft platform state monitoring, equipment parameter interpretation and fault interpretation and have achieved some results, but there is a lack of self-integrated learning satellite anomaly detection methods that can adapt to multi-view and multi-layer feature high-dimensional telemetry and obtain high distinction between normal and abnormal samples. SUMMARY

[0005] The present application aims to overcome the defects of lack of abnormal feature distinction and loss of abnormal decision information in the prior art of directly processing telemetry parameter data, and proposes a self-integrated satellite anomaly detection method and system based on center-constrained contrast learning feature transformation.

[0006] In order to achieve the above-mentioned purpose, the present application proposes a self-integrated satellite anomaly detection method based on center-constrained contrast learning feature transformation, which comprises:

[0007] receiving telemetry raw data downloaded by a spacecraft;

[0008] According to the spacecraft load single machine device to be judged, the telemetry raw data is extracted to obtain corresponding sub-packet telemetry data;

[0009] The pre-processed sub-packet telemetry data is input into a pre-established and trained abnormality discrimination model to realize load single machine device state discrimination;

[0010] The abnormality discrimination model is based on a self-integrated anomaly detection method of center constraint and contrast constraint learning, realizes feature mapping through fusion of center loss and contrast loss, and realizes anomaly detection of spacecraft telemetry data through self-integrated learning of multi-view and multi-level features.

[0011] As an improvement of the above method, the preprocessing includes outlier rejection, missing value filling, state variable one-hot encoding and continuous variable normalization.

[0012] As an improvement of the above method, the sub-packet telemetry parameters x of the spacecraft load single machine device satisfy the following formula:

[0013] x=(d1,...,d c ) k ,...d k )

[0014] Wherein, d θ is the value of the kth parameter; c is the number of parameters.

[0015] As an improvement of the above method, the abnormality discrimination model includes a feature space creation module, a sample feature self-integrated anomaly score function module and an anomaly score state discrimination function module; wherein,

[0016] The feature space creation module includes an encoder f θ , used for extracting data features;

[0017] The sample feature self-integrated anomaly score function module adopts a self-integrated decision method framework, and the feature outputs of each layer are used as inputs of the self-integrated decision SED and output to the anomaly score function;

[0018] The anomaly score state discrimination function module is used for comparing the anomaly score with a threshold to realize anomaly detection.

[0019] As an improvement of the above method, the method further includes a training step of the encoder f θ , specifically including:

[0020] Establish a training set with normal telemetry data;

[0021] SimCLR is used as the contrastive learning framework, and data augmenters are constructed sequentially in conjunction with the central constraint objective. encoder f θ , mapping header g v And the comparison loss function and the center loss function;

[0022] The encoder f is trained according to the set learning rate. θ Training is conducted to obtain a well-trained f θ .

[0023] As an improvement to the above method, the processing procedure of the contrastive learning framework specifically includes:

[0024] Data Augmenter For the input training data x i Perform different random transformations generate and Two augmented samples, and x i First-person and second-person perspectives and For positive sample pairs, and Compared with other samples x j Enhanced samples with j ≠ i form negative sample pairs;

[0025] Among them, data augmentation Using a random affine transformation function:

[0026]

[0027] In the formula, ω and b are the random transformation matrix and random vector sampled from the Gaussian distribution, respectively, and x represents the sample;

[0028] Contrast loss function Each augmented sample of the same sample and As negative sample pairs, different samples x i and x j The same perspective x i (k) and x j (k) As positive sample pairs:

[0029]

[0030] In the formula, It is a set containing K different random affine transformation functions; This will enhance positive samples The sample set after T transformation of the middle sample;

[0031] The and are obtained by f θ (·) respectively, and the corresponding first output and second output

[0032] Through the multi-layer perception network g v (·) mapping h i (1) and h i (2) to a new feature space, obtaining the corresponding mapping z i (1) =g(h i (1) ) and z i (2) =g(h i (2) ), and applying the framework contrast loss function in the space

[0033]

[0034] In the formula, the enhanced positive sample The enhanced negative sample

[0035] Where the enhanced sample contrast loss function is:

[0036]

[0037] In the formula, |{x +}| is the cardinality of the set {x +}; The hyperparameter τ>0 is the temperature coefficient; z(x) is the feature representation of sample x; sim(z,z′)=z T z′ / (||z|||z′||) is the cosine distance between features z and z′;

[0038] The center loss function is obtained according to the following formula is:

[0039]

[0040] In the formula, φ(x i ) is the feature of sample x i ; is the center of the y i th class sample feature, by punishing the distance between the feature and its class center, reducing the intra-class variance;

[0041] is a learnable parameter, which is updated in a mini-batch manner at each iteration;

[0042] The center-constrained contrastive loss function is obtained according to the following formula is:

[0043]

[0044] In the formula, λ is a hyperparameter for balancing the contrastive loss and the center loss.

[0045] As an improvement of the above method, the sample feature self-integrated anomaly score function module uses Mahalanobis distance as an anomaly metric, and a self-integrated decision method SED that fuses multi-view and multi-layer features, including multi-view feature integration MVFE and multi-layer feature integration MLFE, uses information of each neural network layer and each view to improve model discrimination of anomalies, and obtains a sample feature self-integrated anomaly score.

[0046] As an improvement of the above method, the sample feature self-integrated anomaly score satisfies the following formula:

[0047]

[0048] In the formula, s SED (x) is a self-integrated anomaly score of sample x, mean represents a mean aggregation function; L is the number of encoder layers, and n is 1 to L; s mvfe (x, l n ) is a multi-view feature integrated anomaly score of sample x at the l n th layer, and satisfies the following formula:

[0049]

[0050] In the formula, F agg is an aggregation function, and is used to calculate the mean, maximum, or minimum of elements in a set, is a set of anomaly scores of all views of sample x at the l n th layer; K is the number of views, and m is from 1 to K; s m (x, l n ) is an anomaly score of the encoder f θ at the l n th layer, and satisfies the following formula:

[0051]

[0052] In the formula, x (m) is the mth view of sample x, is the l (m) th layer feature of x n , and and respectively, and the mean and covariance matrix of

[0053] In another aspect, the present application also provides a self-integrated satellite anomaly detection system based on center-constrained contrastive learning feature transformation, which comprises a receiving module, an extracting module, a state discrimination module and an anomaly discrimination model, wherein

[0054] The receiving module is configured to receive telemetry raw data transmitted by a spacecraft.

[0055] The extracting module is configured to extract corresponding sub-packet telemetry data from the telemetry raw data according to a spacecraft load single-machine device to be judged.

[0056] The state discrimination module is configured to input the preprocessed sub-packet telemetry data into the pre-established and trained anomaly discrimination model to realize state discrimination of the load single-machine device.

[0057] The anomaly discrimination model is based on a self-integrated anomaly detection method of center-constrained and contrastive learning, realizes feature mapping by fusing center loss and contrast loss, and realizes anomaly detection of the spacecraft telemetry data by using a self-integrated learning method of multi-view and multi-level features.

[0058] Compared with the prior art, the present application has the following advantages:

[0059] 1. The self-integrated satellite anomaly detection method and system based on center-constrained contrastive learning feature transformation provided by the present application can be applied to spacecraft task anomaly detection, and the data-driven method improves the automation level.

[0060] 2. The self-integrated satellite anomaly detection method and system based on center-constrained contrastive learning feature transformation provided by the present application has better algorithm effectiveness and anti-interference than mainstream anomaly detection algorithms, and can provide more effective interpretation support for spacecraft on-orbit operation.

[0061] 3. The center-constrained contrastive learning feature transformation can effectively deal with the case that the normal and abnormal samples of the spacecraft high-dimensional parameters are not obviously distinguished, and the multi-view and multi-level feature self-integrated strategy can effectively overcome the loss of abnormal decision information caused by single-view and single-level features, so that the real state of the spacecraft can be reflected. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 is a center-constrained contrastive learning framework diagram of the present application;

[0063] Figure 2 is a center-constrained self-integrated decision framework diagram of the present application;

[0064] Figure 3 ​is the component schematic diagram of the self-integrated satellite anomaly detection method and system of the center-constrained contrast learning feature transformation of the application;

[0065] Figure 4 is a flowchart schematic diagram of embodiment 1 of the application;

[0066] Figure 5 is a comparison diagram of the anti-interference of the self-integrated satellite anomaly detection method using the center-constrained contrast learning feature transformation of the application and other algorithms to noise pollution, wherein figure 5(a) is the F1 Score value comparison of different algorithms under different pollution rates, and figure 5(b) is the AUC curve comparison under different pollution rates;

[0067] Figure 6 is a comparison diagram of the self-integrated satellite anomaly detection method using the center-constrained contrast learning feature transformation of the application and other algorithms, wherein figure 6(a) is the F1 Score comparison, and figure 6(b) is the AUC curve comparison. DETAILED DESCRIPTION

[0068] The application provides a self-integrated anomaly detection method based on center constraint and contrast constraint learning to solve the problems of lack of distinguishing degree of abnormal features and loss of abnormal decision information in the direct processing of telemetry parameters by conventional anomaly detection algorithms, and maps the normal sample space to a more compact feature distribution space in a manner of fusing center loss and contrast loss, and performs anomaly detection on spacecraft telemetry data samples in a self-integrated learning manner of multi-view and multi-level features.

[0069] The method of the application comprises:

[0070] A telemetry raw data receiving module for receiving satellite downlink telemetry raw data;

[0071] A packet telemetry data extraction module;

[0072] A telemetry parameter formatted dataset construction module for constructing a dataset according to the extracted telemetry data;

[0073] A data preprocessing module, including operations such as outlier rejection, missing value filling, state variable one-hot encoding, and continuous variable normalization;

[0074] A self-integrated anomaly discrimination model of center-constrained contrast learning, including a feature space creation module based on center constraint and contrast constraint, a sample feature self-integrated anomaly score function construction module, and a state discrimination function module for constructing an anomaly score;

[0075] A model parameter tuning and application module.

[0076] To complete the model establishment, the satellite anomaly detection model defines the following related quantities:

[0077] A telemetry parameter vector x, telemetry parameters of a certain subsystem or a certain device.

[0078] x = (d1,...,d k ,...d c ) (1)

[0079] wherein: d k is the value of the kth parameter; c is the number of parameters.

[0080] q is the health state of a certain subsystem or a certain device, q e {0, 1}. q = 0 represents normal, and q = 1 represents abnormal.

[0081] The technical solutions of the present application will be described in detail below in combination with the drawings and examples.

[0082] Example 1

[0083] The embodiment 1 of the present application proposes a self-integrated satellite anomaly detection method based on center-constrained contrastive learning feature transformation, which comprises:

[0084] Receiving telemetry raw data transmitted by a spacecraft;

[0085] According to the spacecraft load single machine device to be judged, the telemetry raw data is extracted to obtain corresponding subpackage telemetry data;

[0086] The preprocessed subpackage telemetry data is input into a pre-established and trained anomaly discrimination model to realize load single machine device state discrimination;

[0087] The anomaly discrimination model is based on a self-integrated anomaly detection method of center-constrained and contrastive constraint learning, realizes feature mapping by fusing center loss and contrast loss, and realizes anomaly detection of spacecraft telemetry data by using a self-integrated learning method of multi-view and multi-level features.

[0088] The self-integrated anomaly discrimination model based on center-constrained contrastive learning is a self-integrated satellite anomaly detection model based on center-constrained contrastive learning feature transformation, which can be formally expressed as:

[0089]

[0090] Wherein, φ is a feature extraction function, is a center-constrained and contrastive constraint feature space construction component module of the model; s is a sample feature construction anomaly score function, is a sample feature self-integrated anomaly score function module based on Mahalanobis distance used by the model; Δ is a discrimination function, is a state discrimination function module of the model using an anomaly score and a threshold. The formal definitions are as follows,

[0091] Let Ω x represent the mThe anomaly detection problem can be described as seeking a target function ψ, which satisfies:

[0092] ψ:Ω x →q (3)

[0093] Given a training sample set For each sample (x (i) ,q (i) ), q (i) = 0. First, learn the feature extraction function φ according to the training sample set, and let Ω h be the feature space of the sample, then the function satisfies:

[0094] φ:Ω x →Ω h (4)

[0095] Sample feature construction anomaly score function s :

[0096] s:Ω h →R (5) In the formula: R is a real number space.

[0097] Anomaly score construction discriminant function Δ:

[0098] Δ:R→q (6)

[0099] The center constraint of the self-integrated anomaly discriminant model of the center constraint contrast learning adopts a center constraint contrast learning framework as shown in Figure 1 , and the self-integrated anomaly score function module of the construction sample feature of the self-integrated anomaly discriminant model of the center constraint contrast learning adopts a self-integrated decision method framework as shown in Figure 2 .

[0100] Figure 1 Among them, the center constraint contrast learning adopts SimCLR as a contrast learning framework, and combines a center constraint target, including a data enhancer , a feature extractor f θ , a mapping head g v , and a contrast loss function and a center loss function.

[0101] Different random transformations are made on the input data x i to generate and two enhanced samples, and respectively called the 1st and 2nd views of x i , and called positive sample pairs,​ and with other samples x j , i≠j form negative sample pairs.

[0102] Data augmenter Random affine transformation function is adopted:

[0103]

[0104] In the formula: ω and b are random transformation matrix and random vector sampled from Gaussian distribution respectively.

[0105] Contrastive loss function Contrast Shifted Instances (CSI) is applied to each augmented sample of the same sample and as negative sample pairs, different samples x i and x j of the same view x i (k) and x j (k) as positive sample pairs:

[0106]

[0107] In the formula: is a set containing K different random affine transformation functions; is a sample set after T transformation of samples in .

[0108] f θ (·) is a neural network-based encoder for extracting features of data samples, and obtained by f θ (·) respectively and

[0109] g v (·) is a multi-layer perceptron network that maps h i (1) and h i (2) to a new feature space to obtain z i (1) =g(h i (1) ) and z i (2) =g(h i (2) ), in which space the contrastive loss function is applied. The contrastive loss function of the SimCLR framework is:

[0110]

[0111] In the formula:

[0112] The contrastive loss function is defined as:

[0113]

[0114] In the formula: |{x +}| represents the set {x + The cardinality of}; the hyperparameter τ>0 is the temperature coefficient; z(x) is the feature representation of sample x; sim(z,z′)=z T z ′ / (||z||||z′||) is the cosine distance between features z and z′.

[0115] The objective of center loss is to reduce the variance of samples of the same class in the feature space, pulling the sample features as close as possible to the center of their class to obtain a more compact feature distribution. The center loss function is defined as follows:

[0116]

[0117] In the formula: φ(x) i ) is the sample x i Features; For the yth i The center of the class sample features. Reduce intra-class variance by penalizing the distance between a feature and its class center.

[0118] For a learnable parameter, updates are performed in mini-batches in each iteration:

[0119]

[0120] In the formula: δ(s) is the indicator function, and the condition is... s It is 1 when the condition is met, and 0 otherwise; the hyperparameter α is the learning rate; ∈ is a small positive value to avoid the denominator being 0.

[0121] Central constraint contrast loss function definition

[0122]

[0123] In the formula: the hyperparameter λ is used to balance the contrast loss and the center loss, and can be set to 10 based on experience.

[0124] Figure 2In the specific implementation, the self-ensemble decision method framework adopts a self-ensemble decision method (SED) that takes Mahalanobis distance as an anomaly metric and fuses multi-view and multi-layer features, uses information of each neural network layer and each view to improve model discrimination of anomalies, and includes multi-view feature ensembled (MVFE) and multi-layer feature ensembled (MLFE). The SED is based on an encoder f θ The feature output of each layer is taken as input of the self-ensemble decision SED and output to an anomaly score function. The encoder f θ The l-th layer feature of the enhanced sample x (m) is calculated as: m

[0125]

[0126] where x (m) is the m-th view of the sample x; f l (x (m) ) is the feature of x (m) at the l-th layer; and are the mean and covariance matrix of x , respectively, where N is the number of training samples.

[0127] The sample x The multi-view feature ensembled anomaly score of the l-th layer of the sample x is:

[0128]

[0129] where F agg is an aggregation function, which can be mean, max and min (respectively calculating the mean, maximum and minimum of elements in the set); S is a set of anomaly scores of all views of the l-th layer of the sample x; K is the number of views, and m is from 1 to K.

[0130] The self-ensemble anomaly score of the sample x is:

[0131]

[0132] where mean is a mean aggregation function; L is the number of layers of the encoder, and n is from 1 to L.

[0133] Embodiment 2

[0134] As Figure 3 ​As shown, the embodiment 2 of the present application provides a self-integrated satellite anomaly detection system based on center-constrained contrastive learning feature transformation, which comprises a module for receiving telemetry raw data transmitted by a satellite; a module for extracting sub-packet telemetry data; a module for constructing a telemetry parameter formatted data set according to the extracted telemetry data; a data preprocessing module, including outlier rejection, missing value filling, state variable one-hot encoding, continuous variable normalization and the like; a self-integrated anomaly discrimination model based on center-constrained contrastive learning, including a feature space creation module based on center-constrained and contrastive constraints, a sample feature self-integrated anomaly score function module, and a state discrimination function module for constructing an anomaly score; and a model parameter optimization and application module.

[0135] The data stored in the satellite telemetry data original database is all original telemetry data transmitted by the satellite.

[0136] The sub-packet telemetry load data extraction module: according to the required instrument equipment, the telemetry data under the corresponding sub-packet telemetry number is selected.

[0137] The telemetry parameter formatted data set construction module for constructing a data set according to the extracted telemetry data: the parameter data obtained by the previous module is processed to construct a telemetry parameter formatted data sample set.

[0138] The data preprocessing module, including outlier rejection, using the 3σ principle to find and eliminate outliers in the telemetry data sample set through first-order data difference; missing value filling, state variable one-hot encoding, continuous variable normalization and the like;

[0139] The self-integrated anomaly discrimination model based on center-constrained contrastive learning includes a feature space creation module, which is constructed according to the center-constrained contrastive learning framework; a sample feature self-integrated anomaly score function module, which is constructed according to the center-constrained self-integrated framework; and a state discrimination function module for constructing an anomaly score.

[0140] The model parameter optimization and application module: the integrated learning model obtained by training is verified in the validation set, the corresponding model parameters are adjusted, and the model is applied to the actual running task for load single machine equipment state discrimination.

[0141] The processing flow of the self-integrated satellite anomaly detection method based on center-constrained contrastive learning feature transformation is as shown in Figure 4 As shown, first, the raw data of the satellite sub-packet telemetry is subjected to preprocessing operations such as outlier rejection, missing value filling and normalization; second, the formatted data set is divided into training data and test data, and the training data only contains normal data samples; third, the training data is used to train the CCL model to obtain the encoder f θ ; fourth, the test data passes through f θ(·) mapped to the feature space, and an anomaly score is obtained by mapping from the ensemble decision function; finally, according to the trained anomaly score and threshold model, it is determined whether the sample is normal.

[0142] Figure 5 is a comparison chart of the anti-interference performance of the self-integrated satellite anomaly detection method using the center-constrained contrastive learning feature transformation of the present application and other algorithms on noise pollution, which describes the average F1 and AUC of CCL-SED and other benchmark algorithms on the Micius dataset with the change of pollution rate. As can be seen from the figure, with the increase of pollution rate, the average F1 and AUC of each algorithm decrease, but the sensitivity is different. The CCL-SED algorithm has a smaller decline rate of F1 and AUC when the pollution rate changes from 0 to 5%, and can maintain the best performance compared with other benchmark algorithms, indicating that the CCL-SED algorithm has good noise anti-interference performance. Among them, Figure 5(a) is the comparison of F1 Score values of different algorithms under different pollution rates, and Figure 5(b) is the comparison of AUC curves under different pollution rates;

[0143] Figure 6 is a comparison chart of the self-integrated satellite anomaly detection method using the center-constrained contrastive learning feature transformation of the present application and other algorithms, which describes the performance improvement of the center loss self-integrated strategy on the CCL-SED algorithm, and shows the mean and standard deviation of F1 and AUC of different combinations of center loss and self-integration on the Micius, Thyroid and Arrhythmia datasets. As can be seen from the figure, without center loss and without using integration, the performance of the algorithm is the worst; while using center loss and self-integration strategy, the algorithm performs best; the improvement brought by center loss and integration strategy is different. Figure 6(a) is the comparison of F1 Score, and Figure 6(b) is the comparison of AUC curve.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent replacements of the technical solutions of the present application do not deviate from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A self-integrated satellite anomaly detection method of center-constrained contrastive learning feature transformation, the method comprising: receiving telemetry raw data transmitted by a spacecraft; extracting the telemetry raw data according to a spacecraft load single machine device to be judged to obtain corresponding sub-package telemetry data; inputting the preprocessed sub-package telemetry data into a pre-established and trained anomaly discrimination model to realize state discrimination of the load single machine device; the anomaly discrimination model is a self-integrated anomaly detection method based on center-constrained and contrastive constraint learning, realizes feature mapping in a manner of fusing center loss and contrastive loss, and realizes anomaly detection of spacecraft telemetry data in a self-integrated learning manner of multi-view and multi-level features; the anomaly discrimination model comprises a feature space creation module, a sample feature self-integrated anomaly score function module, and an anomaly score state discrimination function module; wherein The feature space creating module comprises an encoder f θ for extracting data features; The sample feature self-integrated anomaly score function module adopts a self-integrated decision method framework, and is based on an encoder f θ The feature outputs of the layers are taken as inputs of the self-integrated decision SED and are output to the anomaly score function; the anomaly score state discrimination function module is configured to compare the anomaly score with a threshold to realize anomaly detection; The method further comprises a training step of the encoder f θ , in particular comprising: a training set is established with normal telemetry data; SimCLR is used as a contrast learning framework, and a center constraint target is combined to sequentially construct a data enhancer Encoder f θ , mapping head g v and contrast loss function and center loss function; According to the set learning rate, the encoder f θ is trained to obtain the trained f θ ; the sample feature self-integrated anomaly score function module takes Mahalanobis distance as an anomaly measure, fuses a self-integrated decision method SED of multi-view and multi-level features, includes multi-view feature integration MVFE and multi-level feature integration MLFE, utilizes information of each neural network layer and each view, improves model discrimination of anomalies, and obtains a sample feature self-integrated anomaly score; the sample feature self-integrated anomaly score satisfies the following formula: In the formula, s SED (x) is the self-integrated anomaly score of sample x, mean represents the mean aggregation function; L is the number of encoder layers, n is 1 to L; s mvfe (x, l n ) is the multi-view feature integrated anomaly score of sample x at the l n th layer, which satisfies the following formula: In the formula, F agg These are aggregate functions used to calculate the mean, maximum, or minimum value of elements in a set. For sample x, all viewpoints at the lth n The set of anomaly scores for each layer; K is the number of viewpoints, m ranges from 1 to K; s m (x,l n ) is the encoder f θ In the l n The anomaly score of a layer satisfies the following formula: where x (m) is the mth view of the sample x, is the x (m) lth n feature of the lth layer, and are the mean and covariance matrix of x , respectively, and N is the number of samples in the training set, and T denotes the transpose.

2. The self-integrated satellite anomaly detection method of center-constrained contrastive learning feature transformation according to claim 1, characterized in that, the preprocessing includes outlier rejection, missing value filling, state variable one-hot encoding, and continuous variable normalization.

3. The self-integrated satellite anomaly detection method of center-constrained contrastive learning feature transformation according to claim 1, characterized in that, sub-package telemetry parameters x of a spacecraft load single machine device satisfy the following formula: x = (d1,...,d k ,...d c ) wherein d k is the value of the kth parameter; c is the number of parameters.

4. The self-integrated satellite anomaly detection method of center-constrained contrastive learning feature transformation according to claim 1, characterized in that, the processing process of the contrastive learning framework specifically comprises: Data augmenter On the input training data x i Make different random transformations Generate And Two augmented samples, And The first view and the second view of x i , respectively, And Are positive sample pairs, And With other samples x j , i≠j form negative sample pairs; wherein the data augmenter using a random affine transformation function: in the formula, ω and b are respectively a random transformation matrix and a random vector sampled from a Gaussian distribution, and x represents a sample; Contrastive loss function Each augmented sample of the same sample And Different samples x i And x j The same view x i (k) And x j (k) As a positive sample pair: wherein is a set of K different random affine transformation functions; is a set of positive samples augmented by is a set of samples after T transformation of the samples in will be described below. and by f θ (·) respectively obtain the corresponding first output and second output and a second output through a multi-layer perceptron network g v (·) h i (1) and h i (2) to a new feature space, obtaining the corresponding mapping z i (1) = g(h i (1) ) and z i (2) = g(h i (2) ), in which the framework contrast loss function is applied wherein the enhanced positive samples enhanced negative samples wherein the enhanced sample contrast loss function is: where |{x +} | is the cardinality of the set {x +}; the hyperparameter τ > 0 is a temperature coefficient; z(x) is a feature representation of sample x; sim(z, z') = z T z' / (||z|| ||z' ||) is the cosine distance between features z and z'. The center loss function is obtained according to the following formula is: where φ(x i ) is a feature of sample x i ; is the center of the y i th class sample feature, reduces the intra-class variance by penalizing the distance between a feature and its class center. is a learnable parameter that is updated in small batches each iteration; The center constraint contrast loss function is obtained according to the following formula is: in the formula, λ is a hyperparameter for balancing the contrastive loss and the center loss.

5. A system for self-integrated satellite anomaly detection based on the center-constrained contrastive learning feature transformation of claim 1, the system comprising: a receiving module, an extracting module, a state discrimination module, and an anomaly discrimination model; wherein the receiving module is configured to receive telemetry raw data transmitted by a spacecraft; the extracting module is configured to extract the telemetry raw data according to a spacecraft load single machine device to be judged to obtain corresponding sub-package telemetry data; the state discrimination module is configured to input the preprocessed sub-package telemetry data into a pre-established and trained anomaly discrimination model to realize state discrimination of the load single machine device; the anomaly discrimination model is a self-integrated anomaly detection method based on center-constrained and contrastive constraint learning, realizes feature mapping in a manner of fusing center loss and contrastive loss, and realizes anomaly detection of spacecraft telemetry data in a self-integrated learning manner of multi-view and multi-level features.

Images