Satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement

Abstract

The present invention discloses a satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement, the method comprising: inputting telemetry data of a satellite A and a B payload subsystem for a period of time into a parameter anomaly detection model of a satellite A and a B payload subsystem that has been established and trained in advance to achieve anomaly detection; the parameter anomaly detection model achieves effective fusion of multivariate parameters based on the symbolic aggregation approximation algorithm SAX and the dynamic time warping algorithm Fast‑DTW to form an optimized similarity-based anomaly detection. The present invention solves the problem of information loss caused by not considering context information when fusing multivariate parameters in current anomaly detection methods, achieves effective fusion of multivariate parameters, divides the data into pseudo-periodic sequences with better performance by analyzing the characteristics of satellite telemetry parameter data, and effectively improves anomaly detection results; experiments have shown that the present invention improves the safety of satellite in-orbit operation.

Description

Technical Field

[0001] The present invention belongs to the technical field of satellite anomaly detection, and in particular to a satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement. Background Art

[0002] During the operation of satellites in orbit, due to the harsh space environment, in-orbit abnormal events often occur. In-orbit abnormalities will affect the reliability of satellite in-orbit operation and the safety of space missions to be performed. Satellite engineering parameters are an important basis for ground operation systems to judge the operating status and working performance of satellites. If the abnormalities of parameters can be effectively identified, it is very important to improve the safety and reliability of satellite operation. The purpose of anomaly detection is to monitor the operating status and detect whether the overall operation is normal based on the data of telemetry parameters. In this process, two tasks are mainly completed: one is the selection of telemetry parameters, and the other is anomaly detection. Satellite engineering parameter data has the characteristics of high dimension, time series, pseudo-periodicity, and large data volume. At present, there are insufficient research results on satellite anomaly detection of multi-dimensional parameters, which is one of the important directions for future development.

[0003] At present, the research on satellite anomaly detection at home and abroad can be divided into two categories: methods based on manual combined threshold interpretation, rule-based, physical model-based, and data-driven methods. The first type of research methods have the characteristics of strong interpretability, low false detection rate, and simple implementation. They can effectively detect single-parameter over-limit anomalies and a small number of multi-parameter related anomalies with low complexity. However, these methods usually rely on strict expert experience to formulate inference rules or thresholds, and are powerless for many unknown or undefined anomalies. In addition, the cost of establishing an expert knowledge base and a physical model is high, and for spacecraft systems with large differences, it is often necessary to rebuild the knowledge base and model.

[0004] The data-driven time series anomaly detection method models telemetry data to identify abnormal patterns that are different from normal data. It can be roughly divided into methods based on similarity measurement, machine learning, and state estimation. Some scholars proposed an anomaly detection method based on dynamic time warping (DTW) to calculate the similarity measurement of sequence time with inconsistent lengths; some scholars designed hardware and software based on multi-layer feedforward neural networks to address the problem that spacecraft are susceptible to interference during fault location, and the signal blocking effect is excellent; some scholars proposed a technology based on deep learning and GPU computing to realize spacecraft fault judgment, realize automatic learning and detection of spacecraft faults, and perform well for unknown faults; some scholars used LSSVM to realize anomaly detection of telemetry parameters of satellite power supply components; and some scholars established an ARIMA-SVR combined model to predict the trend of real telemetry data, which improved the prediction accuracy. The data-driven method does not rely on the limitations of expert experience, and can still present good anomaly detection results when expert knowledge is insufficient.

[0005] Summarize and analyze the existing methods related to anomaly detection of satellite multivariate telemetry parameter data:

[0006] (1) The principal component analysis (PCA) method mainly used for dimensionality reduction is a "compression" dimensionality reduction method and is not suitable for dimensionality reduction of data with too high a dimension;

[0007] (2) Feature extraction is often based on classical similarity or distance measurement methods to calculate unit parameters, often ignoring whether multivariate parameters can be effectively integrated;

[0008] (3) Commonly used subsequence segmentation algorithms are not applicable to satellite time series data with typical pseudo-periodicity. Summary of the invention

[0009] The purpose of the present invention is to overcome the defects of the prior art, propose a satellite multi-parameter anomaly detection method based on symbolic representation similarity measurement, based on the symbolic aggregation approximation algorithm (SAX) and the dynamic time warping algorithm (Fast-DTW), study the anomaly detection problem based on multi-satellite engineering time series data, solve the problem of information loss caused by not considering context information when fusing multi-parameters in the current anomaly detection method, realize the effective fusion of multi-parameters, and form an optimized anomaly detection algorithm based on similarity.

[0010] In order to achieve the above object, the present invention proposes a satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement, the method comprising:

[0011] The telemetry data of the A satellite B payload subsystem for a period of time is input into the pre-established and trained parameter anomaly detection model of the A satellite B payload subsystem to realize anomaly detection;

[0012] The parameter anomaly detection model realizes the effective fusion of multivariate parameters based on the symbolic aggregation approximation algorithm SAX and the dynamic time warping algorithm Fast-DTW to form an optimized similarity-based anomaly detection.

[0013] As an improvement of the above method, the method further includes the step of training a parameter anomaly detection model of the A satellite B payload subsystem, specifically including:

[0014] Step s1) performing subsequence segmentation on the time series-based telemetry data of the A satellite B payload subsystem to form a plurality of extreme value-based pseudo-periodic subsequence sets;

[0015] Step s2) selecting m-element parameters of the A satellite B payload subsystem;

[0016] Step s3) performing multi-parameter symbolic representation on the pseudo-periodic subsequence set including the m-element parameter, and performing multi-parameter fusion on the symbolized pseudo-periodic subsequence set by using the N-Gram algorithm in combination with context information;

[0017] Step s4) Based on the Fast-DTW method, similarity measurement is performed on the pseudo-periodic subsequence set after multi-parameter fusion to determine the model parameters of the parameter anomaly detection model, thereby completing the training of the parameter anomaly detection model of the A satellite B payload subsystem.

[0018] As an improvement of the above method, the step s1) specifically includes:

[0019] Step s1-1) according to the preset time initial value and time threshold, find the extreme value points from the continuous telemetry pseudo-periodic sequence to obtain the extreme value point set;

[0020] Step s1-2) According to the preset extreme point initial value and sliding window, the extreme point set is traversed to obtain multiple extreme value-based pseudo-periodic subsequence sets.

[0021] As an improvement of the above method, the step s1-1) specifically includes:

[0022] Step s1-1-1) inputting a continuous telemetry pseudo-period sequence;

[0023] Step s1-1-2) Calculate the sliding window size as ω based on the initial time value t0 and the time threshold threshold i for:

[0024] ω i =t0+threshold

[0025] Step s1-1-3) Find [t0,ω i+t0] corresponds to the time point t of the first extreme point in the range i , the extreme point is a maximum value or a minimum value;

[0026] Step s1-1-4) Find [t i +ω i -threshold,t i +ω i +threshold] The time point corresponding to the extreme point within the period range is t j ;

[0027] Step s1-1-5) Set t i and t j Add to the set of segmentation extreme points and obtain t i and t j The sequence between is added to the pseudo-periodic pattern subsequence set;

[0028] Step s1-1-6) Set t i Assign to t j ;

[0029] Step s1-1-7) If the sequence has not ended, go to step s1-1-3), otherwise obtain a set of extreme points.

[0030] As an improvement of the above method, the step s1-2) specifically includes:

[0031] Step s1-2-1) inputting a set of extreme value points;

[0032] Step s1-2-2) preset the initial value of the extreme point to be e0, and the sliding window size to be len;

[0033] Step s1-2-3) In [e0, e 0+ len] interval to find the first extreme value e i ;

[0034] Step s1-2-4) in [e i , e i+ len] interval to find the second extreme value e j ;

[0035] Step s1-2-5) Determine e i Is it equal to e j , judged as yes, e j Add to the subsequence segment time point set, otherwise e j Assign to e i ;

[0036] Step s1-2-6) If the extreme point set is not completed, go to step s1-2-4); otherwise, go to step s1-2-7);

[0037] Step s1-2-7) Set the sequence between every two time points in the subsequence segmentation time point set as a subsequence, and add it to the pseudo-periodic sequence set, thereby obtaining multiple extreme value-based pseudo-periodic subsequence sets.

[0038] As an improvement of the above method, the m-element parameters of step s2) include 1 telemetry main parameter and m-1 related telemetry parameters.

[0039] As an improvement of the above method, the step s3) specifically includes:

[0040] Step s3-1) using the symbolic aggregation approximation algorithm SAX to symbolize each selected pseudo-periodic subsequence including the m-element parameter, and discretize them into Q strings respectively, forming an m*Q string matrix;

[0041] Step s3-2) Comprehensively represent the multi-dimensional time series context information to obtain the m*3 context information matrix U j :

[0042] U j ={A j-1 A j A j+1}

[0043] Among them, A j-1 , A j and A j+1 Respectively represent the j-1th, jth and j+1th m-ary string sequences after symbolization;

[0044] Step s3-3) uses a 1-Gram representation method to sequentially represent each U from left to right and from top to bottom through a sliding window of size 1. j Divide into several phrases, count the number of times each phrase appears, and then get the frequency to form A j The eigenvector interval U′ j :

[0045] U′ j ={(g1,y1),(g2,y2),…,(g o ,y o )}

[0046] Among them, g o Represents the oth string, y o Represents the string g o Frequency of occurrence;

[0047] Step s3-4) represents the m metadata sequence A with a time series length of n as having a frequency y wOne-dimensional string sequence B, w is the length of string sequence B:

[0048] B:{(B1,y1),(B2,y2),…,(B w ,y w )}.

[0049] As an improvement of the above method, the model parameters of step s4) include the number of letters Sliding window length len, moving step step and truncation number k.

[0050] A satellite multivariate parameter anomaly detection system based on symbolic representation similarity measurement, the system comprising: a detection output module and a parameter anomaly detection model; wherein,

[0051] The detection output module is used to input the telemetry data of the A satellite B payload subsystem for a period of time into a pre-established and trained parameter anomaly detection model of the A satellite B payload subsystem to achieve anomaly detection;

[0052] The parameter anomaly detection model realizes the effective fusion of multivariate parameters based on the symbolic aggregation approximation algorithm SAX and the dynamic time warping algorithm Fast-DTW to form an optimized similarity-based anomaly detection.

[0053] Compared with the prior art, the advantages of the present invention are:

[0054] 1. The present invention uses real satellite telemetry parameter data, based on the symbolic aggregation approximation algorithm (SAX) and the dynamic time warping algorithm (Fast-DTW), to study the anomaly detection problem of multivariate satellite engineering time series data, solve the problem of information loss caused by not considering context information when fusing multivariate parameters in the current anomaly detection method, realize the effective fusion of multivariate parameters, and form an optimized anomaly detection algorithm based on similarity;

[0055] 2. The present invention also proposes a pseudo-periodic time series segmentation algorithm based on extreme values, which divides the data into pseudo-periodic sequences with better performance by analyzing the characteristics of satellite telemetry parameter data, thereby effectively improving the anomaly detection results;

[0056] 3. The research results show that the experimental model obtained by the method of the present invention has good performance in recall, precision and F1 score in the anomaly detection process in the real satellite power subsystem, and has excellent performance in combined anomaly detection. It can be actually applied to satellite anomaly detection to improve the safety of satellite in-orbit operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a flow chart of parameter anomaly detection model training of the present invention;

[0058] Figure 2 This is the schematic result of the segmented sequence when len is 5;

[0059] Figure 3 It is the influence of the number of parameter selections on the evaluation index of anomaly detection results;

[0060] Figure 4 shows the results of various evaluation indicators during the parameter training process, where Figure 4(a) is the result diagram of k=[5,6,7] when len=120 and step=12, Figure 4(b) is the result diagram of len=[60,120,240,480] when k=6 and step=0.1*len, and Figure 4(c) is the result diagram of step=[30,72,144,240] when k=5. DETAILED DESCRIPTION

[0061] Based on the symbolic aggregation estimation algorithm (SAX) and the dynamic time warping algorithm (fast-DTW), the anomaly detection problem of multi-satellite engineering time series data is studied to solve the problem of information loss caused by not considering context information when fusing multi-parameters in the current anomaly detection method, realize the effective fusion of multi-parameters, and form a new anomaly detection algorithm based on similarity;

[0062] The technical process of the satellite anomaly detection algorithm based on EMF (EPA_Segment, MSAX_NG and FastDTW) adopted in the present invention is: firstly, real satellite telemetry data is obtained, the data is preprocessed, and then the processed data is sub-sequenced to form multiple pseudo-periodic sub-sequence sets, and the sub-sequence sets are modeled for anomaly detection and the model parameters are adjusted according to the results of the test set to obtain the best detection model. Finally, the real-time data is used to realize anomaly detection based on the modeling results.

[0063] (1) Pseudo-periodic time series segmentation (EPA_Segment) algorithm based on extreme values

[0064] Satellite time series data generally has the characteristics of high dimension and large data volume. Since the satellite is affected by the sun, orbit, SAA and other factors, it also has typical pseudo-periodicity. The characteristic is that there are a large number of repetitive trends in the parameter data, the overall change of the trend is the same, but there are local differences and the length of the cycle is not the same, and the unchanged central trend cannot be clearly seen. The present invention judges the abnormality of the satellite state based on the similarity measurement, and it is necessary to divide the original parameter data into periodic sequences showing similar rules. The period division process can be regarded as segmenting the parameter sequence into subsequences by finding a certain rule. Satellite time series data has obvious periodicity, and the overall change trend is similar, but there are slight differences in the local area, showing typical pseudo-periodicity. The traditional fixed window subsequence segmentation algorithm for period division cannot solve the problem of satellite parameter data subsequence segmentation well. The length of each subsequence is not the same, but there are often typical jump values ​​in each subsequence, reflecting the overall characteristics of this subsequence. Since the periodic law of satellite parameter data presents an undulating state based on the satellite orbital period, the period division needs to set a sliding window for segmentation with reference to the influence of the orbital period. In summary, according to the characteristics of satellite time series data, a pseudo-periodic time series data subsequence segmentation algorithm based on extreme values ​​is proposed.

[0065] The extreme value-based pseudo-periodic time series segmentation algorithm (EPA_Segment) is divided into two parts: extreme value point acquisition (EPA) and segmentation (Segment).

[0066] The extreme value acquisition (EPA) algorithm can be calculated based on the maximum or minimum value. The algorithm flow is as follows:

[0067] 1) Input a continuous telemetry pseudo-period sequence;

[0068] 2) Calculate the sliding window size as ω i , the initial time value is t0, the time threshold is threshold, then we have

[0069] ω i =t0+threshold

[0070] 3) Find [t0,ω i +t0] the first maximum (minimum) time point t i

[0071] 4) Find [t i +ω i -threshold,t i +ω i +threshold] The time point t of the maximum (minimum) value within a large period j .

[0072] 5) Set ti and t j Add to the set of segmentation extreme points, and obtain the sequence between them, adding pseudo-period

[0073] In the period pattern subsequence set;

[0074] 6) Let t j =t i ;

[0075] 7) If the sequence has not ended, go to step 3, otherwise the EPA algorithm ends.

[0076] The segment algorithm is calculated based on the extreme point set obtained by EPA, and the algorithm flow is as follows.

[0077] 1) Input extreme point set

[0078] 2) Set the initial value of the extreme point to e0 and the sliding window size to len

[0079] 3) In [e0, e 0+ len] interval to find the maximum (minimum) value e i

[0080] 4) In [e i , e i+ Find the maximum (minimum) value e in the interval len] j

[0081] 5) If e i =e j , e j Add to the subsequence segment time point set

[0082] 6) Let e i =e j

[0083] 7) If the extreme point set is not complete, go to 4)

[0084] 8) The sequence between every two time points in the subsequence segmentation time point set is set as a subsequence and added to the pseudo-periodic sequence set.

[0085] 9) The Segment algorithm ends.

[0086] A set of pseudo-periodic time series are generated by the extreme value-based pseudo-periodic time series segmentation (EPA_Segment) algorithm. (2) Multivariate Parameter Symbolic Representation (MASX_NG) algorithm

[0087] Anomaly detection based on multivariate engineering parameters can be classified as anomaly detection of a subsystem. Single parameter anomaly detection is the basis for multivariate parameter detection, and the key to multivariate parameter anomaly detection is the selection and fusion of multivariate parameters. SAX is a classic symbolic representation method that symbolizes single parameter time series data and reduces data dimension. The classic SAX algorithm can only directly symbolize univariate time series data. In addition, when SAX symbolizes data, it only uses the mean to represent the feature, which will cause a relatively large information loss. As a commonly used modeling method in the field of natural processing, the n-gram model (N-Gram) is a probabilistic language model based on a (n-1)-order Markov chain and is widely used in the fields of text classification and speech recognition. This model overcomes the disadvantage of language dependence in conventional algorithms and can extract potential features. The features obtained by this model can improve the efficiency of computer operations and can also be used to extract text features based on the similarity of two strings.

[0088] Based on this, this topic proposes a multi-parameter symbolic representation method based on the classic SAX algorithm for single-parameter symbolic representation, and then fuses the symbolized multi-string set with context information using a sliding window, and then uses the N-Gram algorithm to construct new features.

[0089] In summary, the multivariate parameter symbolic representation algorithm (MSAX_NG) used in this paper is designed as follows:

[0090] 1) First, use the symbolic aggregation approximation algorithm SAX to symbolize the selected m-tuple parameter set

[0091] 2) The N-Gram algorithm is used on the symbolized parameter set to combine the context information for multivariate parameter fusion. The classic SAX algorithm can reduce a time series of any length n to any length w (usually w<<n). It is known that the univariate time series data of length n is T=(t1,t2,....t j ,…,t n ). The steps for symbolic representation using SAX are as follows.

[0092] 1) Subsequence extraction. Use sliding window l and moving step size p to extract the subsequence set of T with length n as follows:

[0093]

[0094] 2) Subsequence standardization. Subsequence S i Standardized to a standard sequence S′ with a mean of 0 and a standard deviation of 1 i .

[0095] S′ i =(t′ (i-1)*p+1 ,t′(i-1)*p+2 ,…,t′ (i-1)*p+l )

[0096] 3) Subsequence dimensionality reduction. The standardized subsequence S′ i Divide into k cutoffs, and take the mean of the parameter data within each cutoff to represent each cutoff.

[0097]

[0098] 4) Subsequence tokenization. Choose an alphabet size of Find the split point β in Table 1 according to the alphabet size i , and then according to Figure 3 The mean Represented by the corresponding letter Each subsequence is tokenized as a string like this

[0099]

[0100] 5) Finally, the time series T is represented as T′.

[0101]

[0102] Given a data sequence A with m-element time series and length n, A=(A1, A2, …, A j ,…,A n ), where A j =(A 1j ,A 2j ,…,A ij ,…,A mj ). The algorithm is designed as follows:

[0103] 1) Discretize each dimensional time series into Q strings according to the SAX algorithm to form an m*Q string matrix.

[0104] 2) In order to minimize the information loss caused by discretization and avoid considering only the information of a single time point while ignoring the contextual information, the multi-dimensional time series contextual information is integrated for representation.

[0105] U j ={A j-1 A j A j+1}

[0106] Among them, A j represents the jth m-element string sequence after symbolization, U j It is an m*3 context information matrix.

[0107] 3) Using the 1-Gram representation method, a sliding window of size 1 is used to sequentially represent each U from left to right and from top to bottom. j Divide it into two parts, each of which becomes an array gram. Count the number of times each gram appears, get the frequency, and form A j The eigenvector interval of .

[0108] U′ j ={(g1,y1),(g2,y2),…,(g o ,y o )}

[0109] Among them, g o Represents the oth string, y o Represents the string g o Frequency of occurrence;

[0110] 4) Thus, the data sequence A with m-tuple time series length n can be represented as a one-dimensional string sequence B with length w and frequency. w is the symbolized string, y k The frequency of occurrence of the string.

[0111] B:{(B1,y1),(B2,y2),…,(B w ,y w )}

[0112] (3) Similarity measurement algorithm

[0113] The dynamic time warping (DTW) algorithm is based on the idea of ​​dynamic programming. In essence, it is to find an optimal path for nonlinear mapping between the template to be matched and the reference template. The biggest advantage of the DTW algorithm is that it can match time series of different lengths and is suitable for satellite engineering parameter data with pseudo-periodic characteristics.

[0114] Suppose there are two time series symbolized string sets A and B, with lengths m and n respectively (m and n can be different), that is

[0115] A:{(A1,x1),(A2,x2),…,(A m ,x m )}

[0116] B:{(B1,y1),(B2,y2),…,(B n ,y n )}

[0117] A1 represents the string, and x1 represents the frequency of the string. If it is a single parameter, the frequency is always 1.

[0118] The DTW algorithm uses dynamic programming to find the shortest path that meets the above conditions in the distance matrix. According to the matching principle, let (p, q) be a point in the matching path. According to the DTW matching path rule, its next point is (p, q+1), (p+1, q) or (p+1, q+1). The point with the smallest distance from point (p, q) to these three points is taken as the next starting point. The distance recursion formula is:

[0119]

[0120] The distance calculation formula used in this article is the MINDIST method used in the SAX algorithm. The calculation formula is:

[0121]

[0122] The DTW algorithm constructs an m*n distance matrix, in which the element d(A i ,B j ) means A i With B j Combined weight x i ,y j The calculated distance between the two.

[0123]

[0124] Let the matching path between A and B be W, w k =(j,i) means that the kth element in path W is A i Map to B j To ensure the validity of the match, the matching path W needs to meet the following three constraints:

[0125] 1) Boundary conditions: w1 = (1,1), w K =(m,n).

[0126] 2) Continuity: Let w k =(p,q),w k+1 =(p′,q′), which must satisfy p′-p≤1 and q′-q≤1.

[0127] 3) Monotonicity: Let w k =(p,q),w k+1 =(p′,q′), p′-p≥0 and q′-q≥0 must be satisfied.

[0128] The DTW algorithm calculates the distance between all components of two time series, but many distance components will not be matched in the matching path process. Fast-DTW is a commonly used acceleration method for DTW, which mainly accelerates the calculation of DTW by adding constraints and data abstraction. Adding constraints reduces the number of searches by constraining the search space. Data abstraction can perform coarse-graining on the original time series iteratively for multiple times, run the DTW algorithm on the time series at a coarser granularity, and then further fine-grained to a finer granularity time series. The coarse-grained data point is the average value of its corresponding multiple fine-grained data points. The strategy of adding constraints is essentially to reduce the search space, so the distance calculated by Fast-DTW is not as accurate as the DTW algorithm, but the time complexity is greatly reduced to O(n). In long time series matching, the amount of calculation and the occupied system space can be greatly reduced during matching, effectively improving the efficiency of anomaly detection. Therefore, the present invention chooses to use the Fast-DTW algorithm for similarity measurement.

[0129] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0130] Example 1

[0131] like Figure 1 As shown, Example 1 of the present invention proposes a satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement, and uses the telemetry parameters of the main power package of a satellite power subsystem as experimental data for modeling and verification. Without defining abnormal labels, the detection ability of the algorithm proposed in the present invention for unknown anomalies is verified, and the detection effect of the algorithm is compared with the currently commonly used supervised learning algorithm KNN and the unsupervised learning algorithm KMeans.

[0132] (1) EPA_Segment algorithm parameter training and experimental results

[0133] The experimental data is divided into periods by EPA_Segment to obtain a pseudo-periodic sequence set. Comprehensive analysis of the algorithm shows that EPA_Segment requires training parameter sliding window len to obtain the optimal segmentation result. The present invention selects {2, 3, 4, 5, 6, 7, 8}, and the training parameters obtain the best segmentation effect.

[0134] The results of the parameter training experiment are shown in Table 1. As can be seen from Table 1, the segmentation effect is best when the sliding window len is 5 or 6 based on the minimum value segmentation calculation, and the segmentation effect is similar when the sliding window len is 6 or 7 based on the maximum value segmentation calculation. This paper selects the result based on the minimum value calculation when the sliding window is 5. The experimental results are shown in Figure 2 It can be clearly seen that the divided sequences present very similar periodic trends.

[0135] Table 1 EPA_Segment algorithm parameter optimization results

[0136]

[0137] (2) Feature selection parameter training results

[0138] The grey correlation analysis method is used for feature selection. First, the main parameters need to be determined. According to experience, in the power supply system, the charging current is selected as the main parameter. The parameter that the algorithm needs to determine is selectparametersnums, that is, the number of parameter selections. The present invention selects the setting as {3,4,5,6,7,8} for training.

[0139] The influence of selectparametersnums on the evaluation index of anomaly detection results is as follows Figure 3 As shown. The Recall index is sensitive to its performance. In the end, simply increasing the number of feature selections will not affect any evaluation index. After comprehensive analysis, selectparametersnums=7 is selected as the optimal feature selection parameter

[0140] The results of the power subsystem feature selection are shown in Table 2. The final data set, together with the main parameters, contains a total of 7 dimensions.

[0141] Table 2 Optimal results of power subsystem feature selection

[0142]

[0143]

[0144] (3) MSAX_NG algorithm parameter training results

[0145] The experimental data training set is used to train the parameters of the MSAX_NG algorithm. Comprehensive analysis algorithm, MSAX_NG algorithm training parameters include the number of letters Sliding window length len, moving step step and truncation number k.

[0146] For the truncation number k, the present invention selects {5, 6, 7}; for the number of letters The present invention selects {3, 4, 5}; for the sliding window len, the present invention selects {60, 120, 240, 480}; for the moving step step, the present invention selects {0.5, 0.7, 0.9} times of the sliding window.

[0147] The training parameters are used to find a set of optimal parameter values. It performs best on the anomaly detection evaluation index, and then directly verifies the model on the validation set.

[0148] The experiment was run on the training set according to the MSAX_NG parameter training process, and the results of each evaluation index are shown in Figure 4.

[0149] As shown in Figure 4(a), the truncation number k increases from 5 to 7, and each indicator shows a trend of first increasing and then decreasing. When k=6, each indicator performs best. As shown in Figure 4(b), the sliding window len doubles from 60 to 480, and each indicator shows a short increase and then a rapid decrease. In particular, recall decreases significantly after 120, while precision increases slowly and then decreases. This is mainly because if the sliding window is too small, the abnormal segment data will be segmented, causing its abnormal feature performance to weaken. If the sliding window is too large, the abnormal segment data and normal data will be mixed too much, resulting in a decrease in recall rate. When len=120, each indicator performs best. As shown in Figure 4(c), the moving step size step increases from 30 to 240. The step size will significantly affect the recall rate, but has no obvious effect on precision. When step=72, each indicator performs best. For the number of letters, From 4 to 5, usually the best case is that the indicators are When Better performance, but It is more susceptible to the influence of sliding window and moving step size, so the stability is not as good as Good performance. In summary, select the collection The model was validated as the optimal parameter set.

[0150] (4) Comparison of algorithms and experimental results

[0151] The present invention selects the supervised learning algorithm 1-NN ED and the unsupervised learning algorithm KMeans ED for comparison. The results of the comparison algorithms are shown in Table 3.

[0152] Table 3 Anomaly detection results of different algorithms

[0153]

[0154] Through statistics of the three evaluation indicators of each algorithm, the results show that the EMF algorithm used in the present invention is better than the 1-NN ED and KMeans ED algorithms in all indicators and performs best. Comprehensive evaluation shows that the basic order of performance is EMF>1-NNED>KMeans ED.

[0155] KMeans ED performs worse than 1-NN ED, but it is an unsupervised learning algorithm and is more suitable for detecting unknown anomalies. The precision of 1-NN ED and EMF are similar, but the recall of EMF is significantly better than 1-NN ED, indicating that it has a more comprehensive ability to detect anomalies in the anomaly detection process.

[0156] Example 2

[0157] Embodiment 2 of the present invention proposes a satellite multivariate parameter anomaly detection system based on symbolic representation similarity measurement, which is implemented based on the method of embodiment 1. The system includes: a detection output module and a parameter anomaly detection model; wherein,

[0158] The detection output module is used to input the telemetry data of the A satellite B payload subsystem for a period of time into a pre-established and trained parameter anomaly detection model of the A satellite B payload subsystem to achieve anomaly detection;

[0159] The parameter anomaly detection model realizes the effective fusion of multivariate parameters based on the symbolic aggregation approximation algorithm SAX and the dynamic time warping algorithm Fast-DTW to form an optimized similarity-based anomaly detection.

[0160] Innovation

[0161] (1) The present invention proposes a period division method suitable for pseudo-periodic satellite time series data

[0162] (2) The present invention proposes a symbolic representation method for feature fusion for multivariate parameters

[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. A satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement, the method comprising: The telemetry data of the A satellite B payload subsystem for a period of time is input into the pre-established and trained parameter anomaly detection model of the A satellite B payload subsystem to realize anomaly detection; The parameter anomaly detection model realizes effective fusion of multivariate parameters based on the symbolic aggregation approximation algorithm SAX and the dynamic time warping algorithm Fast-DTW to form an optimized similarity-based anomaly detection; The method further includes the step of training a parameter anomaly detection model for the A satellite B payload subsystem, specifically including: Step s1) performing subsequence segmentation on the time series-based telemetry data of the A satellite B payload subsystem to form a plurality of extreme value-based pseudo-periodic subsequence sets; Step s2) selecting m-element parameters of the A satellite B payload subsystem; Step s3) performing multi-parameter symbolic representation on the pseudo-periodic subsequence set including the m-element parameter, and performing multi-parameter fusion on the symbolized pseudo-periodic subsequence set by using the N-Gram algorithm in combination with context information; Step s4) Based on the Fast-DTW method, similarity measurement is performed on the pseudo-periodic subsequence set after multivariate parameter fusion to determine the model parameters of the parameter anomaly detection model, thereby completing the training of the parameter anomaly detection model of the A satellite B payload subsystem; The step s3) specifically includes: Step s3-1) using the symbolic aggregation approximation algorithm SAX to symbolize each selected pseudo-periodic subsequence including the m-element parameter, and discretize them into Q strings respectively, forming an m*Q string matrix; Step s3-2) Comprehensively represent the multi-dimensional time series context information to obtain the m*3 context information matrix U j : U j ={A j-1 A j A j+1 } Among them, A j-1 , A j and A j+1 Respectively represent the j-1th, jth and j+1th m-ary string sequences after symbolization; Step s3-3) uses a 1-Gram representation method to sequentially represent each U from left to right and from top to bottom through a sliding window of size 1. j Divide into several phrases, count the number of times each phrase appears, and then get the frequency to form A j The eigenvector interval U′ j : U′ j ={(g1,y1),(g2,y2),…,(g o ,y o )} Among them, g o Represents the oth string, y o Represents the string g o Frequency of occurrence; Step s3-4) represents the m metadata sequence A with a time series length of n as having a frequency y w One-dimensional string sequence B, w is the length of string sequence B: <h2 style=";text-align:left;direction:ltr">B:{(B1,y1),(B2,y2),…,(B<h2 style=";text-align:left;direction:ltr"> w <h2 style=";text-align:left;direction:ltr"> ,y<h2 style=";text-align:left;direction:ltr"> w <h2 style=";text-align:left;direction:ltr"> )}。 2. The satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement according to claim 1, characterized in that: The step s1) specifically comprises: Step s1-1) according to the preset time initial value and time threshold, find the extreme value points from the continuous telemetry pseudo-periodic sequence to obtain the extreme value point set; Step s1-2) According to the preset extreme point initial value and sliding window, the extreme point set is traversed to obtain multiple extreme value-based pseudo-periodic subsequence sets.

3. The satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement according to claim 2, characterized in that: The step s1-1) specifically includes: Step s1-1-1) inputting a continuous telemetry pseudo-period sequence; Step s1-1-2) Calculate the sliding window size as ω based on the initial time value t0 and the time threshold threshold i for: ω i =t0+threshold Step s1-1-3) Find [t0,ω i +t0] corresponds to the time point t of the first extreme point in the range i , the extreme point is a maximum value or a minimum value; Step s1-1-4) Find [t i +ω i -threshold,t i +ω i +threshold] The time point corresponding to the extreme point within the period range is t j ; Step s1-1-5) Set t i and t j Add to the set of segmentation extreme points and obtain t i and t j The sequence between is added to the pseudo-periodic pattern subsequence set; Step s1-1-6) Set t i Assign to t j ; Step s1-1-7) If the sequence has not ended, go to step s1-1-3), otherwise obtain a set of extreme points.

4. The satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement according to claim 2, characterized in that: The step s1-2) specifically includes: Step s1-2-1) inputting a set of extreme points; Step s1-2-2) preset the initial value of the extreme point to be e0, and the sliding window size to be len; Step s1-2-3) in [e 0， e 0+ len] interval to find the first extreme value e i ; Step s1-2-4) in [e i， e i+ len] interval to find the second extreme value e j ; Step s1-2-5) Determine e i Is it equal to e j , judged as yes, e j Add to the subsequence segment time point set, otherwise e j Assign to e i ; Step s1-2-6) If the extreme point set is not completed, go to step s1-2-4); otherwise, go to step s1-2-7); Step s1-2-7) Set the sequence between every two time points in the subsequence segmentation time point set as a subsequence, and add it to the pseudo-periodic sequence set, thereby obtaining multiple extreme value-based pseudo-periodic subsequence sets.

5. The satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement according to claim 1, characterized in that: The m-element parameters of step s2) include 1 telemetry main parameter and m-1 related telemetry parameters.

6. The satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement according to claim 1, characterized in that: The model parameters of step s4) include the number of letters Sliding window length len, moving step step and truncation number k.

7. A system based on the satellite multivariate parameter anomaly detection method based on symbolic representation similarity measurement according to claim 1, characterized in that: The system comprises: a detection output module and a parameter anomaly detection model; wherein, The detection output module is used to input the telemetry data of the A satellite B payload subsystem for a period of time into a pre-established and trained parameter anomaly detection model of the A satellite B payload subsystem to achieve anomaly detection; The parameter anomaly detection model realizes the effective fusion of multivariate parameters based on the symbolic aggregation approximation algorithm SAX and the dynamic time warping algorithm Fast-DTW to form an optimized similarity-based anomaly detection.

Images