Satellite telemetry data set anomaly detection method based on long time sequence characteristics
Through the Informer model and residual statistics method, the problem of short-term and long-term data detection in satellite telemetry data is solved, the accuracy and stability of abnormal detection are improved, and effective abnormal detection of satellite telemetry data is realized.
Patent Information
- Application Number
- CN202411027883.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-08
AI Technical Summary
Existing satellite telemetry data anomaly detection methods are difficult to take into account short-term and long-term data, especially the mining ability of long-term scale changes, and it is difficult to judge set anomalies by a separate residual threshold method.
Using satellite telemetry data set anomaly detection method based on long-term sequence characteristics, data prediction is carried out by building an Informer model, three anomaly detection thresholds are extracted, and abnormal thresholds are generated using residual statistical distribution, residual mean value and mean square error to achieve iterative prediction. To improve detection accuracy.
It improves the long-term change law capture capability of satellite telemetry data, improves the accuracy of abnormal detection and reduces the error detection rate, and solves the problem of long-term data detection.
Smart Images

Figure CN118916811A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a satellite telemetry data set anomaly detection method based on long time series characteristics, belonging to the technical field of satellite telemetry data anomaly detection. Background Art
[0002] With the rapid development and transformation of my country's aerospace industry from "experimental" to "business", "rapid development and long-term service" of satellites has become the mainstream trend. Correspondingly, condition monitoring and health management technology have become the key supporting technologies for satellite on-orbit maintenance, dynamic mission planning, on-orbit life extension and reliability improvement. In recent years, the number of satellite launches in my country has increased year by year. The high frequency of launches means that the time for satellite design and testing is limited, which increases the risk of satellite on-orbit operation. Ground management is needed to cope with such challenges. In this case, traditional manual interpretation methods will be difficult to ensure the accuracy of satellite anomaly detection and cannot meet the needs of satellite telemetry data mining because they require a large number of human resources for long-term and high-intensity work. In addition, satellite telemetry data is still the only reference for the ground to judge the on-orbit operation status of satellites. With the increase in the number of satellites, the amount of telemetry data that needs to be processed will also increase exponentially. More efficient condition monitoring capabilities are needed to detect anomalies in the data and ensure the safety and reliability of satellite operation.
[0003] Aggregate anomalies refer to the fact that some data points in a data set are not abnormal in themselves, but appear abnormal as a set. Since aggregate anomalies involve joint anomalies of multiple points, single-point detection methods are difficult to effectively identify. This requires the model to be able to capture the inherent correlation characteristics of long-term change trends in order to accurately detect and identify them. In actual working conditions, satellites are affected by seasonal fluctuations in temperature and solar activity cycles, and telemetry data has obvious stability and periodic fragment characteristics over a long period of time. Existing models such as isolation forests and autoencoders often extract the time series evolution laws between dozens of points, that is, the input and output samples contain insufficient time information, and cannot fully reflect the evolution process and trend changes of the system, and it is difficult to capture the long-term change laws of data fragments. In addition, if the change trend of a section of abnormal data is different from that of normal data, but any point inside it is within the normal value range, then using the residual of the observed value and the predicted value as the anomaly judgment criterion will not be able to determine whether the data segment is abnormal. Therefore, the challenges facing current research are as follows:
[0004] (1) Existing anomaly detection methods have difficulty in taking into account both short-term satellite telemetry data and long-period data, especially the ability to mine long-term change trends. A long-term characteristic prediction method is needed.
[0005] (2) When performing anomaly detection on telemetry data, the residual threshold method alone is difficult to determine the aggregate anomaly, and the anomaly judgment criteria need to be reconsidered. Summary of the invention
[0006] The present invention aims to solve the technical problems that the existing satellite telemetry data anomaly detection methods are difficult to take into account both short-time series satellite telemetry data and long-period data, especially the mining ability of long-time scale change trends, and the anomaly judgment criteria need to be recalculated during each detection. A satellite telemetry data set anomaly detection method based on long-time series characteristics is proposed.
[0007] The technical solution adopted by the present invention to solve the above problems is: the present invention proposes a satellite telemetry data set anomaly detection method based on long-term series characteristics, comprising:
[0008] Step 1: Build a satellite telemetry data prediction model;
[0009] Step 2: Preprocess the historical satellite telemetry data;
[0010] Step 3: Training the satellite telemetry data prediction model based on the preprocessed historical satellite telemetry data;
[0011] Step 4: Input the satellite telemetry data samples without abnormalities into the trained satellite telemetry data prediction model to extract three abnormality detection thresholds;
[0012] Step 5: Use the test data input. If the prediction result of a single satellite telemetry data exceeds the threshold, replace the input value of the next round of prediction with the prediction output value of this round, and mark the data abnormal interval. If it does not exceed the threshold, use the satellite telemetry data sample to continue the next round of prediction.
[0013] Step 6: Repeat step 5 to achieve anomaly detection under the premise of long-term prediction of satellite telemetry data.
[0014] Optionally, the input of the satellite telemetry data prediction model in step 1 is the satellite telemetry data matrix X = {x0, x1, x2, ..., x n}, the output is the predicted value of satellite telemetry data, where x i is the telemetry data value obtained at a single time, and the subscripts 0, 1, 2, ..., n are the positions of the corresponding telemetry data in the input sequence;
[0015] The satellite telemetry data prediction model includes an encoder, a decoder, and multiple attention layers;
[0016] The encoder is used to perform pooling operations on the satellite telemetry data after passing through each attention layer;
[0017] The decoder and attention layers are used to compute the satellite telemetry data prediction results of the satellite telemetry data prediction model.
[0018] The satellite telemetry data prediction model used in the present invention is the Informer model. The Informer model is aimed at long-time series prediction. It can not only capture the long-distance dependencies of the input sequence, but also increase the probability attention mechanism to reduce the computational complexity compared to the traditional Transformer model, realize the sequence input of hundreds of points, and facilitate the learning of the intrinsic correlation of long-time scale characteristic sequences.
[0019] Optionally, the step of preprocessing the historical satellite telemetry data in step 2 includes:
[0020] Step 2.1: Assign position codes to the input data of the satellite telemetry data prediction model according to odd and even bits, and store the sequence position information;
[0021] Step 2.2: Convert the timestamp label vector X1 of the input data into a one-dimensional convolutional layer, and return the output result through the transposition operation to complete the preprocessing of the historical satellite telemetry data;
[0022] The expression for assigning positional encoding to input position is:
[0023]
[0024] Formula (1) is the position encoding method of the input data of odd digits, and formula (2) is the position encoding method of the input data of even digits. In formulas (1) and (2), pos is the position vector, i is the index of the data, and the value range is 0≤2i<d, and d is the set model dimension constant;
[0025] The output expression is:
[0026]
[0027] In formula (3), variables i and j are the element positions of the output tensor matrix, and the shape of the input tensor X1 is (N,C in ,L in ), the shape of the convolution kernel W is (C out ,C in ,K), the shape of the output tensor is (N,C out ,L out ), N is the batch size, C out is the number of output channels, b j is the bias term of the jth output channel, s is the stride of the convolution operation, K is the convolution kernel size, and b j is the bias term of the jth output channel;
[0028] The expression for the length of the output sequence is:
[0029]
[0030] In formula (4), L out is the output sequence length, K is the convolution kernel size, L in is the length of the input sequence.
[0031] Optionally, the step of training the satellite telemetry data prediction model in step 3 includes:
[0032] Step 3.1: Assign query matrix Q, key matrix K and value matrix V to the preprocessed historical satellite telemetry data;
[0033] Step 3.2: Query the correlation between Q and K. For each query vector q i , select the partial key vector K of the key matrix K sample Perform inner product calculation to obtain the attention score of the query vector;
[0034] Step 3.3: According to the attention score of the query vector, select U high-scoring items with scores higher than the preset value and calculate the attention weights;
[0035] Step 3.4: Weight the value matrix V based on the calculated attention weights to obtain the output of the attention mechanism, completing the construction of a single attention layer in the satellite telemetry data prediction model;
[0036] Step 3.5: Repeat steps 3.1 to 3.4 to build multiple attention layers. Between adjacent attention layers, the encoder performs a pooling operation to discard half of the data with low weight values. After processing through multiple attention layers, the final output of the encoder is used as K and V, and the output of the previous attention layer of the decoder is used as Q. The last attention mechanism of the decoder is completed, and the output vector Y1 is obtained;
[0037] Step 3.6: Perform a linear transformation on the output vector Y1 to obtain the final output result Y2, completing the training of the satellite telemetry data prediction model;
[0038] The calculation formula of attention weight is:
[0039]
[0040] The expression of the weighted value matrix V is:
[0041]
[0042] In formulas (6) and (7), z i is the i-th element in vector Z, d k is the dimension of the matrix, v j is the jth element in the value matrix V;
[0043] The calculation formula for the final output result Y2 is:
[0044] Y2=Y1W+b(8)
[0045] In formula (8), W is the weight matrix and b is the bias vector.
[0046] Optionally, the step of extracting the data anomaly threshold in step 4 includes:
[0047] Step 4.1: Output of a single prediction of the model and the non-anomalous observation value Y of the corresponding prediction result i =(y i1 ,y i2 ,...,y in ) to obtain the residual matrix E ki , where i is the number of predictions and n is the number of points in a single prediction;
[0048] Step 4.2: Calculate the residual matrix E ki The residual mean and residual standard deviation of the data in the calculation matrix E is calculated based on the mean standard deviation. ki The normal distribution of the data in the calculation is used to fit the residual distribution, and the abnormal threshold of the residual mean is calculated by the normal distribution.
[0049] Step 4.3: Calculate the mean square error threshold based on the normal distribution, and use the mean square error threshold to deal with deviations and outliers in the distribution of satellite telemetry prediction data;
[0050] Step 4.4: Calculate the mean of the residual matrix of satellite telemetry data, the mean μ and standard deviation σ of the normal distribution and mean square error, and set the threshold K of the residual distribution according to the residual mean μ and standard deviation σ;
[0051] The residual matrix E ki The calculation formula is:
[0052]
[0053] The residual matrix E ki The formula for calculating the data mean is:
[0054]
[0055] In formula (11), μ is the mean of the normal distribution;
[0056] The residual matrix E ki The formula for calculating the standard deviation of the data mean is:
[0057]
[0058] In formula (12), σ 2 is the variance;
[0059] The residual matrix E ki The expression for the normal distribution of the data in is:
[0060]
[0061] The formula for calculating the mean square error is:
[0062]
[0063] In formula (13), y i is the observed value, is the predicted value;
[0064] The calculation formula of threshold K is:
[0065]
[0066] Optionally, marking the abnormal data interval in step 5 specifically includes:
[0067] The abnormal threshold of the residual mean, the mean square error threshold and the threshold K of the residual distribution are used in turn to detect the outliers of the satellite telemetry prediction data and complete the calibration of the data anomaly interval.
[0068] Optionally, when the prediction result of a single satellite telemetry data in step 5 exceeds the threshold, the expression of the input value of the next round of prediction is:
[0069]
[0070] When the prediction result of a single satellite telemetry data does not exceed the threshold, the expression of the input value of the next round of prediction is:
[0071]
[0072] In formula (16), y k is the observed value in round i, is the predicted value of the i-th round, d i+1 Input parameters for round i+1.
[0073] The beneficial effects of the present invention are:
[0074] 1. The present invention utilizes a set anomaly detection method optimized by long-time scale characteristic modeling to establish a model to receive longer time series inputs, learn the intrinsic correlation of long-time scale characteristic sequences, capture the long-term change patterns of data segments, and improve the accuracy of prediction.
[0075] 2. The present invention adopts a method based on residual statistical distribution, residual mean, and residual mean square error to generate a satellite telemetry data anomaly detection threshold, and determines the secondary round input of iterative prediction based on a single detection result, thereby avoiding the influence of collective anomalies on the ability to capture long-term scale characteristics, thereby improving the accuracy of anomaly detection and reducing the false detection rate, and solving the problem that current anomaly detection methods and judgment criteria are difficult to take into account long-period data. BRIEF DESCRIPTION OF THE DRAWINGS
[0076] Figure 1 A flowchart of a satellite telemetry data set anomaly detection method based on long-time series features provided by the present invention;
[0077] Figure 2 A schematic diagram of the satellite telemetry data prediction model provided by the present invention;
[0078] Figure 3 The error feature extraction flow chart provided by the present invention;
[0079] Figure 4 A sample diagram of the roll axis attitude angle provided by the present invention;
[0080] Figure 5 A sample diagram of the yaw axis attitude angle provided by the present invention;
[0081] Figure 6 The anomaly detection result diagram based on mean square error provided by the present invention;
[0082] Figure 7 This is a satellite telemetry data anomaly interval annotation diagram provided by the present invention. DETAILED DESCRIPTION
[0083] Specific implementation method 1: Combination Figure 1-2 To illustrate this embodiment, the steps of a satellite telemetry data set anomaly detection method based on long-term series features described in this embodiment include:
[0084] S1: Build satellite telemetry data prediction model;
[0085] S101: Figure 2 As shown in the figure, the input of the satellite telemetry data prediction model is the satellite telemetry data matrix X = {x0, x1, x2, ..., x n}, the output is the predicted value of satellite telemetry data, where x i is the telemetry data value obtained at a single time, and the subscripts 0, 1, 2, ..., n are the positions of the corresponding telemetry data in the input sequence;
[0086] The satellite telemetry data prediction model includes an encoder, a decoder, and multiple attention layers;
[0087] The encoder is used to perform pooling operations on the satellite telemetry data after passing through each attention layer;
[0088] The decoder and attention layers are used to calculate the satellite telemetry data prediction results of the satellite telemetry data prediction model;
[0089] The satellite telemetry data model constructed by the present invention is an Informer model. The satellite telemetry data model is aimed at long-time series prediction. It can not only capture the long-distance dependencies of the input sequence, but also increase the probabilistic attention mechanism to reduce the computational complexity compared to the traditional Transformer model, realize the sequence input of hundreds of points, and facilitate the learning of the intrinsic correlation of long-time scale characteristic sequences.
[0090] S102: Separate the timestamp and value of the input data, assign position codes to the input data according to odd and even bits according to formula (1) and formula (2), store the sequence position information, and input it into the attention mechanism.
[0091]
[0092] Formula (1) is the position encoding method for odd-numbered input data, and formula (2) is the position encoding method for even-numbered input data. In formulas (1) and (2), pos is the position vector, i is the index of the data, and its value range is 0≤2i<d, and d is the set model dimension constant.
[0093] S2: Data preprocessing: Convert the timestamp label vector X1 of the data into a one-dimensional convolutional layer and return the output result through the transpose operation.
[0094] The output expression is:
[0095]
[0096] In formula (3), variables i and j refer to the element positions of the output tensor matrix, and the shape of the input tensor X1 is (N,C in ,L in ), the shape of the convolution kernel W is (C out ,C in ,K), the shape of the output tensor is (N,C out ,L out ), where N is the batch size, C out is the number of output channels, b j represents the bias term of the jth output channel, s is the stride of the convolution operation, K is the convolution kernel size, and b j represents the bias term of the jth output channel;
[0097] The expression for the length of the output sequence is:
[0098]
[0099] In formula (4), L out is the output sequence length, K is the convolution kernel size, L in The length of the input sequence in bits.
[0100] S3: training satellite telemetry data model;
[0101] S301: Assign the query matrix Q (query), key matrix K (key) and value matrix V (value) to the numerical features of the output results. Query the degree of association between Q and K. For each query vector q i , select the partial key vector K of the key matrix K sample Perform inner product calculation to form an attention score;
[0102] The calculation formula of attention weight is:
[0103]
[0104] S302: According to the obtained attention scores, select U high-scoring items with higher scores, then calculate the attention weights, and perform weighted summation on the value matrix V as shown in formula (6) and formula (7). Based on the calculated attention weights, weight V to obtain the output of the attention mechanism, thus completing the construction of an attention layer.
[0105] The expression of the weighted value matrix V is:
[0106]
[0107] In formulas (6) and (7), z i is the i-th element in vector Z, d k is the dimension of the matrix, v j is the jth element in the value matrix V;
[0108] S303: Repeat S301-S302 to complete the construction of multiple attention layers of the satellite telemetry data model. Between the attention layers, the encoder performs a pooling operation to discard half of the data with lower weights, and then passes through the next attention layer. Finally, after multiple attention layers are processed, the final output of the encoder is used as K and V, and the output of the previous attention layer of the decoder is used as Q. The last attention mechanism of the decoder is completed. The obtained output vector Y1 (attitude angle tensor) is linearly transformed by formula (8) to become the final output result, completing the model training. The shape of the tensor Y1 in formula (8) is the same as X1 in formula (3), and the shapes of the weight matrix W and the bias vector b are also the same as formula (3).
[0109] The calculation formula for the final output result Y2 is:
[0110] Y2=Y1W+b(8)
[0111] In formula (8), W is the weight matrix and b is the bias vector.
[0112] S4: Input the telemetry data sample into the satellite telemetry data model, and judge whether the single prediction result exceeds the threshold according to the abnormal threshold of the residual mean, the mean square error threshold and the threshold K of the residual distribution. If there is an abnormality in the current round of data, replace the input value of the next round of prediction with the predicted output value of this round, and mark the abnormal interval. When the detection result is normal, the input value of the next round of prediction is shown in formula (9). If there is no abnormality, the observed data is used to continue the next round of prediction, and the input value is shown in formula (10);
[0113] When the prediction result of a single satellite telemetry data exceeds the threshold, the expression of the input value of the next round of prediction is:
[0114]
[0115] When the prediction result of a single satellite telemetry data does not exceed the threshold, the expression of the input value of the next round of prediction is:
[0116]
[0117] In formula (10), y k is the observed value in round i, is the predicted value of the i-th round, d i+1 Input parameters for round i+1.
[0118] Finally, the satellite telemetry data set anomaly detection method based on long time series characteristics proposed in this embodiment can eliminate the interference of abnormal data and ensure the stability of prediction under the premise of long time series prediction of satellite telemetry data, and can locate the abnormal interval and realize the detection of abnormal data.
[0119] Specific implementation method 2: Figure 3 As shown, the step of error feature extraction described in this embodiment includes:
[0120] S5: Output of a single prediction of the satellite telemetry data prediction model Where i is the number of predictions, and n is the number of points in a single prediction. The corresponding value is the non-abnormal observation value Y of the prediction result. i =(y i1 ,y i2 ,...,y in ), the difference between the two is used to obtain the final residual matrix E ki ;
[0121] The residual matrix E ki The calculation formula is:
[0122]
[0123] S6: Fit the probability distribution according to the data of the residual matrix, and judge the trend through the mean and standard deviation. Select the normal distribution to fit the residual distribution, and its probability density function is as shown in formula (12). Use the maximum likelihood estimation to calculate the model parameters so that the probability of the observed data is maximized. The solution process is shown in formula (13) and formula (14);
[0124] The residual matrix E ki The formula for calculating the data mean is:
[0125]
[0126] In formula (12), μ is the mean of the normal distribution;
[0127] The residual matrix E ki The formula for calculating the standard deviation of the data mean is:
[0128]
[0129] In formula (13), σ 2 is the variance;
[0130] The residual matrix E ki The expression for the normal distribution of the data in is:
[0131]
[0132] S7: If there are large deviations or outliers in the distribution of the prediction results, it may not accurately reflect the accuracy of the prediction. Therefore, the mean square error threshold processing is used, such as formula (15), where y i is the observed value, is the predicted value.
[0133] The formula for calculating the mean square error is:
[0134]
[0135] In formula (15), y i is the observed value, is the predicted value;
[0136] S8: The prediction residual obeys the normal distribution. The residual mean and standard deviation of each time point are calculated using formula (12) and formula (13). The threshold K of the residual distribution is set according to the mean μ and standard deviation σ to detect abnormal points and complete the abnormal threshold extraction.
[0137] The calculation formula of threshold K is:
[0138]
[0139] To summarize, this implementation adopts a method based on residual statistical distribution, residual mean, and residual mean square error to generate a satellite telemetry data anomaly detection threshold, and determines the secondary round input of iterative prediction based on a single detection result, thereby avoiding the impact of collective anomalies on the ability to capture long-term scale characteristics, thereby improving the accuracy of anomaly detection and reducing the false detection rate, and solving the problem that current anomaly detection methods and judgment criteria are difficult to take into account long-period data.
[0140] Example
[0141] Combination Figure 4-7 This embodiment is described. This embodiment uses the attitude angle telemetry data of the real satellite attitude and orbit control system to verify the method. Figure 4 and Figure 5 As shown, Figure 4 The figure shows the changes in the roll axis attitude angle from 7:56 on February 16, 2011 to 8:55 on February 17, 2011. The samples are taken in units of minutes, with about 1500 sampling points. It can be seen that the roll axis attitude angle cycle is about 12 hours. Figure 5 This is the yaw axis attitude angle sample, which shows the changes in the yaw axis attitude angle from 9:23 on April 26, 2011 to 10:22 on April 27. The sample is also sampled in minutes, with about 1500 sampling points. Unlike the roll angle sample, the sample period is about 6 hours.
[0142] The satellite telemetry data prediction model is used to predict the long-term characteristics of the yaw axis and roll axis attitude angle data. The residual mean and mean square error of the prediction results are calculated using formulas (12) and (15). The 24-point and 48-point inputs are used for comparative analysis to verify the long-term characteristic relationship capture capability and model stability of the method. The telemetry data prediction results are shown in Table 1 below:
[0143] Table 1
[0144]
[0145] After completing the long-term characteristic prediction, the anomaly threshold corresponding to the residual mean, mean square error and residual distribution is calculated by formula (10) as the basis for anomaly detection. The anomaly detection threshold is as follows: Figure 2 As shown:
[0146] Table 2
[0147]
[0148] After determining the threshold, iterative prediction is performed and the single prediction result is tested. If any of the three indicators does not meet the requirements, it is judged as abnormal data and the interval will be marked as an abnormal interval. The detection results based on the mean square error and the abnormal interval marking are as follows: Figure 6 and Figure 7 shown.
[0149] Based on the above method, anomaly detection is performed on the roll axis attitude angle and yaw axis attitude angle samples. The 24-point prediction, 48-point prediction and isolation forest method are used for comparison. The comparison results of anomaly detection are shown in Table 3:
[0150] Table 3
[0151]
[0152] It can be seen from Table 3 that the prediction accuracy of the satellite telemetry data set anomaly detection method based on long time series characteristics proposed in this embodiment is much greater than that of the isolation forest method. Therefore, the present invention can eliminate the interference of abnormal data and ensure the stability of the prediction under the premise of long time series prediction of satellite telemetry data, and can locate the abnormal interval to realize the detection of abnormal data.
[0153] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.
Claims
1. A satellite telemetry data set anomaly detection method based on long time series characteristics, characterized in that: The method for detecting anomaly in satellite telemetry data sets based on long-term series features comprises the following steps: Step 1: Build a satellite telemetry data prediction model; Step 2: Preprocess the historical satellite telemetry data; Step 3: Training the satellite telemetry data prediction model based on the preprocessed historical satellite telemetry data; Step 4: Input the satellite telemetry data samples without abnormalities into the trained satellite telemetry data prediction model to extract three abnormality detection thresholds; Step 5: Use the test data input. If the prediction result of a single satellite telemetry data exceeds the threshold, replace the input value of the next round of prediction with the prediction output value of this round, and mark the data abnormal interval. If it does not exceed the threshold, use the satellite telemetry data sample to continue the next round of prediction. Step 6: Repeat step 5 to achieve anomaly detection under the premise of long-term prediction of satellite telemetry data.
2. According to claim 1, a satellite telemetry data set anomaly detection method based on long time series features is characterized in that: The input of the satellite telemetry data prediction model in step 1 is the satellite telemetry data matrix X = {x0, x1, x2, ..., x n }, the output is the predicted value of satellite telemetry data, where x i is the telemetry data value obtained at a single time, and the subscripts 0, 1, 2, ..., n are the positions of the corresponding telemetry data in the input sequence; The satellite telemetry data prediction model includes an encoder, a decoder, and multiple attention layers; The encoder is used to perform a pooling operation on the satellite telemetry data after passing through each attention layer; The decoder and attention layer are used to calculate the satellite telemetry data prediction results of the satellite telemetry data prediction model.
3. The method for detecting anomalies in satellite telemetry data sets based on long-term series characteristics according to claim 1, characterized in that: The steps of preprocessing the historical satellite telemetry data in step 2 include: Step 2.1: assigning position codes to the input data of the satellite telemetry data prediction model according to odd and even bits, and storing sequence position information; Step 2.2: Convert the timestamp label vector X1 of the input data into a one-dimensional convolutional layer, and return the output result through the transposition operation to complete the preprocessing of the historical satellite telemetry data; The expression for assigning positional encoding to input position is: Formula (1) is the position encoding method of the input data of odd digits, and formula (2) is the position encoding method of the input data of even digits. In formulas (1) and (2), pos is the position vector, i is the index of the data, and the value range is 0≤2i<d, and d is the set model dimension constant; The output expression is: In formula (3), variables i and j are the element positions of the output tensor matrix, and the shape of the input tensor X1 is (N,C in ,L in ), the shape of the convolution kernel W is (C out ,C in ,K), the shape of the output tensor is (N,C out ,L out ), N is the batch size, C out is the number of output channels, b j is the bias term of the jth output channel, s is the stride of the convolution operation, K is the convolution kernel size, and b j is the bias term of the jth output channel; The expression for the length of the output sequence is: In formula (4), L out is the output sequence length, K is the convolution kernel size, L in is the length of the input sequence.
4. The method for detecting anomalies in satellite telemetry data sets based on long-term series characteristics according to claim 1, characterized in that: The steps of training the satellite telemetry data prediction model in step 3 include: Step 3.1: assigning a query matrix Q, a key matrix K and a value matrix V to the preprocessed historical satellite telemetry data; Step 3.2: Query the correlation between Q and K. For each query vector q i , select the partial key vector K of the key matrix K sample Perform inner product calculation to obtain the attention score of the query vector; Step 3.3: According to the attention score of the query vector, select U high-scoring items with scores higher than a preset value and calculate the attention weights; Step 3.4: Weight the value matrix V based on the calculated attention weights to obtain the output of the attention mechanism, completing the construction of a single attention layer in the satellite telemetry data prediction model; Step 3.5: Repeat steps 3.1 to 3.4 to build multiple attention layers. Between adjacent attention layers, the encoder performs a pooling operation to discard half of the data with low weight values. After processing through multiple attention layers, the final output of the encoder is used as K and V, and the output of the previous attention layer of the decoder is used as Q. The last attention mechanism of the decoder is completed, and the output vector Y1 is obtained; Step 3.6: Perform a linear transformation on the output vector Y1 to obtain the final output result Y2, completing the training of the satellite telemetry data prediction model; The calculation formula of attention weight is: The expression of the weighted value matrix V is: In formulas (6) and (7), z i is the i-th element in vector Z, d k is the dimension of the matrix, v j is the jth element in the value matrix V; The calculation formula for the final output result Y2 is: Y2=Y1W+b(8) In formula (8), W is the weight matrix and b is the bias vector.
5. The method for detecting anomalies in satellite telemetry data sets based on long-term series characteristics according to claim 1, characterized in that: The steps of extracting data anomaly threshold in step 4 include: Step 4.1: Output of a single prediction of the model and the non-anomalous observation value Y of the corresponding prediction result i =(y i1 ,y i2 ,...,y in ) to obtain the residual matrix E ki , where i is the number of predictions and n is the number of points in a single prediction; Step 4.2: Calculate the residual matrix E ki The residual mean and residual standard deviation of the data in the calculation matrix E is calculated based on the mean standard deviation. ki The normal distribution of the data in the calculation is used to fit the residual distribution, and the abnormal threshold of the residual mean is calculated by the normal distribution. Step 4.3: Calculate the mean square error threshold based on the normal distribution, and extract the mean square error threshold to deal with the deviation and outliers in the distribution of satellite telemetry prediction data; Step 4.4: Calculate the mean of the residual matrix of satellite telemetry data, the mean μ and standard deviation σ of the normal distribution and mean square error, and set the threshold K of the residual distribution according to the residual mean μ and standard deviation σ; The residual matrix E ki The calculation formula is: The residual matrix E ki The formula for calculating the data mean is: In formula (11), μ is the mean of the normal distribution; The residual matrix E ki The formula for calculating the standard deviation of the data mean is: In formula (12), σ 2 is the variance; The residual matrix E ki The expression for the normal distribution of the data in is: The formula for calculating the mean square error is: In formula (13), y i is the observed value, is the predicted value; The calculation formula of threshold K is:
6. The method for detecting anomalies in satellite telemetry data sets based on long-term series characteristics according to claim 1, characterized in that: The abnormal intervals of marked data in step 5 specifically include: The abnormal threshold of the residual mean, the mean square error threshold and the threshold K of the residual distribution are used in turn to detect the outliers of the satellite telemetry prediction data and complete the calibration of the data anomaly interval.
7. The method for detecting anomalies in satellite telemetry data sets based on long-term series characteristics according to claim 1, characterized in that: When the prediction result of a single satellite telemetry data in step 5 exceeds the threshold, the expression of the input value of the next round of prediction is: When the prediction result of a single satellite telemetry data does not exceed the threshold, the expression of the input value of the next round of prediction is: In formula (16), y k is the observed value in round i, is the predicted value of the i-th round, d i+1 Input parameters for round i+1.
Citation Information
Patent Citations
Single-dimensional telemetry data anomaly detection method based on ELM
CN109034191A
Informer-based autonomous satellite positioning precision prediction method and device
CN116819581A
BiLSTM-based Internet of Things time series data anomaly detection method
CN117312769A
Facilitating efficient and effective anomaly detection via minimal human interaction
US20220292074A1
