CN109934337B - Spacecraft telemetry data anomaly detection method based on integrated LSTM - Google Patents
Spacecraft telemetry data anomaly detection method based on integrated LSTM Download PDFInfo
- Publication number
- CN109934337B CN109934337B CN201910194332.0A CN201910194332A CN109934337B CN 109934337 B CN109934337 B CN 109934337B CN 201910194332 A CN201910194332 A CN 201910194332A CN 109934337 B CN109934337 B CN 109934337B
- Authority
- CN
- China
- Prior art keywords
- value
- matrix
- lstm model
- telemetering
- remote control
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Landscapes
- Telephonic Communication Services (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
A spacecraft telemetry data anomaly detection method based on integrated LSTM belongs to the technical field of telemetry data anomaly detection. Hair brushThe method obviously solves the problem that the existing method has low accuracy in detecting the context abnormality of the telemetering data. The invention preprocesses the training data, divides the training data into two categories of a training set A and a training set B, and then trains the training set A and the training set B respectively based on the LSTM model, thereby effectively reducing the influence of long-term dependence on the detection result and improving the accuracy of the detection of the whole abnormity; integrating the prediction results of the LSTM model according to a certain weight to obtain the final predicted value of the telemetering data, performing difference by using the error between the prediction result and the actual value, namely, smoothing the error, and performing dynamic threshold value processing according to the errorAnd detecting an abnormal interval of the telemetering data, compared with the prior method, the method can improve the accuracy of the detection of the abnormal context of the telemetering data of the spacecraft by more than 11%. The method can be applied to the technical field of detection of telemetering data abnormity.
Description
Technical Field
The invention belongs to the technical field of detection of telemetering data abnormity, and particularly relates to a spacecraft telemetering data abnormity detection method based on integrated LSTM.
Background
With the gradual increase of the complexity and functionality of spacecraft design, the probability of spacecraft failure caused by various reasons such as insufficient design, manufacturing errors, maintenance errors and unplanned events increases, and as for 764 spacecraft launched from 1993 to 2012 and including satellites and space stations, the spacecraft has 121 failures in total, and accounts for 15.8% of the total number of launched spacecraft. Meanwhile, the research and development period of the satellite is gradually shortened, the emission number is increased, and the strict requirements of safe and reliable operation of the satellite on data processing and analysis cannot be met only by the traditional reliability engineering method and means and the manual analysis method depending on the personal experience of experts.
At present, most satellites are controlled in a ground telemetering and remote control mode, during the in-orbit operation period of the satellites, a ground monitoring station needs to receive and process telemetering data of the in-orbit satellites, automatic judgment, alarming and other operations are carried out according to state monitoring parameters of the in-orbit satellites, and meanwhile, the obtained telemetering data are stored in a database for post data analysis. As the only information source for ground personnel to judge the working state of the orbiting satellite, whether the telemetering data is abnormal or not is closely related to whether the working mode of the satellite is changed or whether the satellite has a fault or not. Therefore, how to pre-judge the time, possibility and severity of the abnormal state by analyzing the satellite in-orbit running state data in real time and accurately, and performing effective abnormal processing and health management according to the pre-judging result becomes a key research for enhancing the satellite in-orbit running reliability.
The research on the anomaly detection of the telemetering data of the spacecraft has attracted extensive attention of domestic and foreign research institutions. For example, the European Space Operations control center (ESOC) affiliated to the ESA has developed an Analysis and Reporting systems (ARES) System, which is the latest off-line data Analysis System solution of the ESOC; ORCA and Induction Monitoring System (IMS) are developed by NASA Emms Research Center (ARC), wherein ORCA excavates abnormalities and outliers by calculating the distance between adjacent points, and IMS software automatically constructs a knowledge base for System health Monitoring; an automatic Telemetry data Health Monitoring System (Automated Telemetry Health Monitoring System, ATHMoS) is developed by the German Space Operation Center (GSOC), and anomaly detection is realized by using outlier detection and supervised and learning methods.
The current telemetering data abnormity detection method is mainly divided into three methods: statistical-based methods, distance-based and prediction-based methods.
Statistical methods typically assume that data obeys a known distribution model, and data that does not conform to the distribution or that the statistical features do not conform are flagged as anomalous, however, actual data is difficult to characterize with a preset distribution. The distance-based method labels the data points with longer distance and lower density by measuring the distance between different data. The method is based on the density, and is assisted by distance measurement, so that the time-series high-accuracy abnormity detection is realized; the literature (research on the similarity measurement method of the time sequence of the satellite telemetry data) develops the research on the similarity measurement method for the time sequence of the satellite telemetry data, and realizes the elimination of the correlation influence among parameters and the realization of asynchronous measurement. However, distance-based methods are very sensitive to distance metric functions and difficult to model temporal correlations. The deviation based method is characterized in that the deviation between the measured data and the model output is counted by modeling the normal data, and the data exceeding the fixed deviation is marked as abnormal. The method based on the deviation has strong interpretability based on a prediction method, strong online application capability and high test speed, and is more suitable for realizing online monitoring of telemetering data.
However, the above researches are based on the unit and multivariate abnormality detection of the telemetry data, and for the actual complex spacecraft system, the ground can send corresponding remote control instructions through a remote control link to effectively control the system. Telemetry and telemetry interact, and it is clear that telemetry without telemetry is blind and telemetry without telemetry is passive. Therefore, how to effectively fuse the remote control instructions to realize the anomaly detection of the telemetry data is an effective way for improving the anomaly detection of the telemetry data.
Recently, as deep learning progresses, many models are applied to the application of anomaly detection. Particularly, the LSTM method, as a special type of Recurrent Neural Network (RNN), retains the advantages of the standard RNN, can perform deduction prediction on a time series by using historical data, and is widely concerned in time series anomaly detection. In the literature (Detecting space analytes Using LSTMs and Nonparametric Dynamic threshold), NASA combines the powerful nonlinear modeling capability and the automatic feature extraction capability of the LSTM, and simultaneously takes a remote control instruction and telemetric data as input to construct an LSTM model, thereby realizing effective marking of the telemetric data. However, because a uniform LSTM model is constructed for the control modes of all instructions, the average value and variance of the overall prediction error are increased, the detection capability of the method for detecting the context anomaly of the spacecraft telemetry data is poor, and the detection accuracy of the context anomaly is low.
Disclosure of Invention
The method aims to solve the problem that the accuracy of detecting the context abnormality of the spacecraft telemetry data is low in the existing method.
The technical scheme adopted by the invention for solving the technical problems is as follows: a spacecraft telemetry data anomaly detection method based on integrated LSTM comprises the following steps:
step one, selecting training dataWherein:is a group of m-dimensional vectors represented by a remote control command matrix at the time t and a remote measuring value at the time t, and the remote control command matrix at the time t istelemetry value at time t ofPreprocessing the selected training data S to obtain reconstructed input X:
step two, respectively extracting a remote control instruction matrix in each secondary matrix contained in the reconstructed input X, wherein the remote control instruction matrix in each secondary matrix is respectively expressed asRespectively calculating the L2 norm of the remote control command matrix corresponding to each secondary matrix, taking the median of all the calculated L2 norms as a threshold beta, taking a set formed by the secondary matrixes corresponding to the remote control command matrix with the L2 norm being greater than the threshold beta as a training set A, and taking a set formed by the secondary matrixes corresponding to the remote control command matrix with the L2 norm being less than or equal to the threshold beta as a training set B;
step three, constructing an integrated LSTM model containing an LSTM model A and an LSTM model B, and respectively training the LSTM model A and the LSTM model B by utilizing the training set A and the training set B obtained in the step two to obtain a well-trained LSTM model A and an LSTM model B;
step four, test data are preprocessed to obtain a test set, the test set is respectively input into the well-trained LSTM model A and the well-trained LSTM model B, and a telemetering predicted value sequence formed by predicted values of the telemetering values corresponding to each secondary matrix output by the LSTM model A is respectively obtainedTelemetering predicted value sequence formed by predicted values of telemetering values corresponding to each secondary matrix output by LSTM model B
Setting fusion parameters, and determining a final telemetering value prediction sequence according to a telemetering prediction value sequence consisting of prediction values of telemetering values corresponding to each secondary matrix output by the LSTM model A, a telemetering prediction value sequence consisting of prediction values of telemetering values corresponding to each secondary matrix output by the LSTM model B and the fusion parameters
Step five, the final telemetering value prediction sequence determined in the step fourThe value in (d) is subtracted from the actual remote value at the corresponding time to obtain an error sequence e ═ e { (e)(1),e(2),…,e(t),…},e(t)Represents the error at time t;
smoothing the error sequence by adopting an exponential weighted average method to obtain a smoothed error sequence
Will smooth out the error sequenceThe smooth error values are grouped, and then the dynamic threshold value of each group is respectively calculatedExceeding dynamic threshold value in each groupAnd the value of the smoothing error is less than the dynamic threshold value in the groupAnd is closest to the dynamic thresholdIf the difference value is greater than the discrimination threshold value p, the dynamic threshold value is exceededThe telemetering value at the moment corresponding to the smoothing error value is regarded as an abnormal value; if the difference is less than or equal to the discrimination threshold p, the exceeding dynamic threshold is determinedThe telemetering value of the moment corresponding to the smoothing error value is taken as a normal value;
and step six, repeating the process from the step four to the step five for the telemetering data of the spacecraft to be detected, and realizing the detection of the abnormity of the telemetering data of the spacecraft to be detected.
The invention has the beneficial effects that: the invention discloses a spacecraft telemetering data anomaly detection method based on integrated LSTM, which comprises the steps of preprocessing training data, dividing the training data into a training set A and a training set B, and then training the training set A and the training set B respectively based on an LSTM model, so that the influence of long-term dependence on a detection result can be effectively reduced, and the accuracy of overall anomaly detection is improved; integrating the prediction results of the LSTM model according to a certain weight to obtain the final predicted value of the telemetering data, performing difference by using the error between the prediction result and the actual value, namely, smoothing the error, and performing dynamic threshold value processing according to the errorAnd detecting an abnormal interval of the telemetering data, compared with the prior method, the method can improve the accuracy of the detection of the abnormal context of the telemetering data of the spacecraft by more than 11%.
Drawings
FIG. 1 is a flow chart of a method of the present invention for integrated LSTM based detection of spacecraft telemetry data anomalies;
FIG. 2 is a schematic diagram of a synchronous many-to-many time series model employed by the present invention;
FIG. 3 is a block diagram of LSTM model prediction;
FIG. 4 is a flow chart of the processing and classification of training data;
FIG. 5 is a comparison graph of an actual telemetry value and a predicted telemetry value;
FIG. 6 is a schematic diagram illustrating the abnormal point determination according to the present invention;
FIG. 7 is a schematic representation of all telemetry values on a telemetry channel;
FIG. 8a) is a schematic diagram of an actual exception sequence;
fig. 8b) is a schematic diagram of the anomaly detection result.
Detailed Description
The first embodiment is as follows: as shown in fig. 1, the method for detecting an anomaly of spacecraft telemetry data based on integrated LSTM according to the embodiment includes the following steps:
step one, selecting training dataWherein:is a group of m-dimensional vectors represented by a remote control command at the time t and a remote measuring value at the time t, and the remote control command at the time t istelemetry value at time t ofPreprocessing the selected training data S to obtain reconstructed input X:
step two, respectively extracting the remote control instruction in each secondary matrix contained in the reconstructed input X, wherein the remote control instruction in each secondary matrix is respectively expressed asRespectively calculating the L2 norm of the remote control command corresponding to each secondary matrix, taking the median of all the calculated L2 norms as a threshold beta, taking a set formed by the secondary matrixes corresponding to the remote control commands of which the L2 norms are greater than the threshold beta as a training set A, and taking a set formed by the secondary matrixes corresponding to the remote control commands of which the L2 norms are less than or equal to the threshold beta as a training set B;
FIG. 4 is a flow chart of the processing and classification of training data;
step three, constructing an integrated LSTM model containing an LSTM model A and an LSTM model B, and respectively training the LSTM model A and the LSTM model B by utilizing the training set A and the training set B obtained in the step two to obtain a well-trained LSTM model A and an LSTM model B;
step four, test data are preprocessed to obtain a test set (the preprocessing of the test data refers to preprocessing to obtain reconstructed input, all secondary matrixes in the reconstructed input are used as the test set), the test set is respectively input into the trained LSTM model A and the trained LSTM model B, and a telemetering predicted value sequence formed by the predicted values of the telemetering data corresponding to each secondary matrix output by the LSTM model A and a telemetering predicted value sequence formed by the predicted values of the telemetering data corresponding to each secondary matrix output by the LSTM model B are respectively obtained;
setting fusion parameters, and determining a final telemetering data prediction sequence according to a telemetering prediction value sequence formed by the prediction values of the telemetering data corresponding to each secondary matrix output by the LSTM model A, a telemetering prediction value sequence formed by the prediction values of the telemetering data corresponding to each secondary matrix output by the LSTM model B and the fusion parameters
Step five, the final telemetering data prediction sequence determined in the step four is usedThe value in (d) is subtracted from the actual remote value at the corresponding time to obtain an error sequence e ═ e { (e)(1),e(2),…,e(t),…},e(t)Represents the error at time t;
smoothing the error sequence by adopting an exponential weighted average (EWMA) method to obtain a smoothed error sequence
Grouping the smooth error values in the smooth error sequence, and respectively calculating the dynamic threshold of each groupExceeding dynamic threshold value in each groupAnd the value of the smoothing error is less than the dynamic threshold value in the groupAnd is closest to the dynamic thresholdIf the difference value is greater than the discrimination threshold value p, the dynamic threshold value is exceededThe telemetering value at the moment corresponding to the smoothing error value is regarded as an abnormal value; if the difference is less than or equal to the discrimination threshold p, the exceeding dynamic threshold is determinedThe telemetering value of the moment corresponding to the smoothing error value is taken as a normal value;
in obtaining aAfter the group has smoothed the error, time 1 to ltTaking the smoothing error in the first group as the smoothing error in the next h time in the first group and the smoothing error in the time lt+1 to 2ltAnd taking the smoothing errors in the error group as a second group, and repeating the steps to complete the grouping of the smoothing errors, and if the quantity of the smoothing errors in the last group is less than h, fusing the smoothing errors with the previous group.
Calculating the mean and variance of each group of smoothing errors, and further calculating a dynamic threshold
Each ltAnd regarding the smoothing errors in the time as a group, and keeping the smoothing errors in the h time in the previous group, so that the smoothing error sequence is divided into a plurality of parts to respectively calculate the dynamic threshold.
smoothing errors for the nth setThe average value of (a) of (b),is composed ofZ is a coefficient matrix.
Since the abnormality occurs at the time when the smoothing error abruptly changes, the points exceeding the threshold cannot be collectively regarded as the abnormal points. Exceeding dynamic threshold within the same groupIs compared with the maximum smoothing error that does not exceed the dynamic threshold, and if the difference between the two points is less than or equal to p, it is considered as a normal point. The principle is as shown in figure 6 of the drawings,andall exceed the threshold value, butAndis less than p (p is 0.01), soIt is still the normal point, the same principleThe abnormality is considered.
And step six, repeating the process from the step four to the step five for the telemetering data of the spacecraft to be detected, and realizing the detection of the abnormity of the telemetering data of the spacecraft to be detected.
The training data and the test data of the present embodiment are derived from the data set collected in the incorporated surservices, analog (isa) reports published by NASA, which relates to two spacecraft types, including smap (soil motion Active and passive) and msl (mars Science laboratory).
The second embodiment is as follows: the first difference between the present embodiment and the specific embodiment is: the specific process of preprocessing the selected training data S to obtain the reconstructed input X is as follows:
splitting the training data S into a plurality of sub-matrices comprising successive time vectors, said sub-matrices being represented respectively as:and the number of columns of each secondary matrix is ltThen the reconstructed input X is represented as:
wherein:the first secondary matrix is represented as a function of,a second sub-matrix is represented which is,representing the nth secondary matrix;
the third concrete implementation mode: the second embodiment is different from the first embodiment in that: the specific process of the second step is as follows:
respectively extracting a remote control instruction matrix in each secondary matrix contained in the reconstructed input X, and respectively representing the remote control instruction matrix in each secondary matrix asWherein:is composed ofThe matrix of remote control commands in (1),is composed ofThe matrix of remote control commands in (1),is composed ofThe matrix of remote controls of (1) is,
wherein:representsThe transpose of (a) is performed,representsThe L2 norm;representing and calculating matrixThe maximum eigenvalue of (d);
after the L2 norm of each remote control command matrix is calculated, the median of all L2 norms is taken as a threshold value beta, a set formed by secondary matrixes corresponding to the remote control command matrix with the L2 norm being less than or equal to the threshold value beta is taken as a training set B, and a set formed by the secondary matrixes corresponding to the remote control command matrix with the L2 norm being greater than the threshold value beta is taken as a training set A.
The fourth concrete implementation mode: the first difference between the present embodiment and the specific embodiment is: the LSTM model A in the third step comprises an input layer, a first hidden layer, a second hidden layer and an output layer, wherein a loss function of the LSTM model A adopts a Mean absolute error function (MAE), an optimization function adopts an adam (adaptive motion estimation) function, and the value of drop is [0.2, 0.3 ];
the structure and parameter settings of the LSTM model B are the same as those of the LSTM model A.
The structures of the LSTM model a and the LSTM model B adopted in the present embodiment are divided into three layers: an Input Layer (Input Layer), a Hidden Layer (Hidden Layer) and an Output Layer (Output Layer), wherein the Hidden Layer is two layers. Each hidden layer contains a plurality of Memory blocks (Memory blocks), and each Memory Block contains a plurality of Memory cells (Memory cells). For the anomaly detection problem to be solved by the invention, the model type adopted is a synchronous many-to-many time series model, as shown in fig. 2, each test data is input and passes through a hidden layer, and then a prediction matrix is synchronously output.
The memory cells in a single memory block are connected together by a chain structure. Based on standard RNN, a memory Cell in LSTM contains one Cell (Cell) and three gates (Gate). The information in the cell is called the cell state, which is updated based on historical output as well as real-time input. The three gates are respectively: an Input Gate (Input Gate), a forgetting Gate (Forget Gate), and an Output Gate (Output Gate). Because the telemetering data is a time sequence, the change of the value of the telemetering data is related to the remote control command at the same time, and during the learning process of the LSTM on the training data, the accuracy of the anomaly detection system can be better improved by selectively adopting the characteristic data.
The training process of the data in the LSTM system is as follows: input data x at time ttAll cells in the same Block as the previous time,as input into the input gate and the forgetting gate. The output of the input gate is:
whereinFor the output of the input gate, f (x) is the Sigmoid function. OmegailFor the ith input at the same timeWeight matrix of ωclFor the output of the c-th cell in the same blockThe weight matrix of (2).
The output of Forget Gate is:
whereinFor the output of the forgetting gate, f (x) is the Sigmoid function.For the ith input at the same timeThe weight matrix of (a) is determined,for the output of the c-th cell in the same blockThe weight matrix of (2).
The output of the cell will combine the output of the input gate and the output of the forgetting gate:
being cellsOutput, g (x), is the tahn function.Update information for candidate cells, ωicFor the ith input at the same timeThe weight matrix of (2).
Final intracellular outputInput x that will correspond to the current timetThe corresponding output is the final prediction result of the current learning as the input of the output gate.
For Output Gate: omegaiωFor the ith input at the same timeWeight matrix of ωiωFor the ith input at the same timeA weight matrix of (a);
ωcωis the output of the c Cell of the same BlockThe weight matrix of (2).For Output Gate Output, f (x) is the activation function.
The learning rate of each parameter is dynamically adjusted by using the first Moment estimation and the second Moment estimation of the gradient due to a model optimization function adam (adaptive motion estimation), and the learning rate has the capability of storing the average value of the previous partial derivative square attenuation.
Through multiple iterative learning, the model can provide higher performance by updating the optimized weight matrix, and in order to prevent overfitting, a certain proportion of memory modules are prevented from participating in the learning process of the LSTM in each hidden layer, which is also called dropout.
The fifth concrete implementation mode: the fourth difference between this embodiment and the specific embodiment is that: in the third step, the training set A and the training set B obtained in the second step are used for respectively training the LSTM model A and the LSTM model B to obtain a trained LSTM model A and a trained LSTM model B; the specific process comprises the following steps:
respectively inputting the training set A and the training set B into an LSTM model A and an LSTM model B for training, setting the maximum iteration times of the LSTM model A and the LSTM model B for training to be alpha, continuously adjusting each weight matrix of the LSTM model A and the LSTM model B in each iteration training process, and stopping training until the maximum iteration times are reached to obtain the well-trained LSTM model A and the LSTM model B.
For any one secondary matrix in input XFor each oneTo say that ltThe sequence of how many consecutive times to determine is used as a set of input values for the LSTM model, and each set of input values is used to predict the telemetry value at the next time.
If the secondary matrixBelonging to the training set A, the secondary matrixInputting LSTM model A for training, and determining the matrixBelonging to training set B, the secondary matrixInputting an LSTM model B for training;
obtaining a predicted telemetry value of the next moment of each iterationAnd analyzing the error of the predicted remote measurement value at the next moment by using the loss function, optimizing the model by using the optimization function based on the result of the loss function until the training is finished when the maximum iteration times are reached, and obtaining a trained LSTM model A and an LSTM model B.
in order to train the weight and find the corresponding gradient, a proper loss function and an optimization function are selected to help the model to determine the corresponding optimal weight matrix, so that the prediction accuracy is improved, and the difference between the true value and the prediction value is reduced. The specific LSTM model framework is shown in fig. 3: in the training process, the training data is automatically adjusted through the numerical value of the optimization function in a multi-iteration mode until the optimal weight is obtained, and the accurate prediction of the time sequence data is realized. After the model training is finished, the model can quickly predict the multidimensional data, and the predicted value is used as an important basis for the abnormal detection of the telemetering data.
The sixth specific implementation mode: the fourth difference between this embodiment and the specific embodiment is that: the specific form of the absolute value error function is as follows:
The seventh embodiment: the first difference between the present embodiment and the specific embodiment is: the telemetry prediction value sequence formed by the prediction values of the telemetry values corresponding to each secondary matrix output by the LSTM model A, the telemetry prediction value sequence formed by the prediction values of the telemetry values corresponding to each secondary matrix output by the LSTM model B and the fusion parameter determine the final telemetry value prediction sequenceThe method comprises the following steps:
for secondary matrix in test setSecondary matrixThe corresponding remote control command matrix isSecondary matrix of the imageInputting the LSTM model A to obtain the predicted value of the remote measurement value output by the LSTM model ASecondary matrix of the imageInputting the LSTM model B to obtain the predicted value of the remote measurement value output by the LSTM model B
Similarly, the predicted value of the remote measurement value output by the LSTM model A and the predicted value of the remote measurement value output by the LSTM model B are input to other secondary matrixes in the test set;
then the remote measuring predicted value sequence formed by the predicted values of the remote measuring values corresponding to each secondary matrix output by the LSTM model AExpressed as:herein, theThe predicted value of telemetry data corresponding to the first submatrix of the test set,for the predicted value of telemetry data corresponding to the second submatrix of the test set,is the t-th sub-matrix of the test set (i.e., the) A predicted value of the corresponding telemetry data;
then the remote measuring predicted value sequence formed by the predicted values of the remote measuring values corresponding to each secondary matrix output by the LSTM model BExpressed as:
if the remote control instruction matrixL2 norm is greater than threshold β, then secondary matrixCorresponding final telemetry value prediction valuesExpressed as:
if the remote control instruction matrixL2 norm is less than or equal to threshold beta, then secondary matrixCorresponding final telemetry value prediction valuesExpressed as:
The specific implementation mode is eight: the first difference between the present embodiment and the specific embodiment is: smoothing the error sequence by adopting an exponential weighted average methodProcessing to obtain a smoothing error sequence es(ii) a The specific process comprises the following steps:
for the smoothed error value at time t,a smoothing error value at time t-1 (when t is 1, the error at that time is a default smoothing error value), and γ is a threshold value of the smoothing error;
As shown in fig. 5, the variation of the error is very large under normal conditions, which is not beneficial to analysis, and most of the error is considered to be caused by the difference between the predicted value and the actual value rather than the occurrence of abnormality. Smoothing the error becomes particularly important. The error e is smoothed, so that an error peak which often occurs when a predicted value is obtained based on the LSTM prediction model can be suppressed. Since the actual telemetry value may be abruptly changed due to the remote command at that time, the abrupt change is usually not well predicted, which will result in an abrupt change of e at a certain time, and the abrupt change is a normal phenomenon and cannot be regarded as abnormal.
In the embodiment, the error value is processed by using a Weighted moving average method (EWMA) to obtain a smooth error, thereby avoiding the sudden change of the error value caused by a large difference between a predicted value and a normal value. And the meaning of judging whether the data is abnormal is given only when the error value after smoothing is abnormally high or abnormally low in a period of time.
Examples
The training data and test data included 82 sets of data out of twelve subsystems, with a total of 55 sets (67%) belonging to SMAP and 27 sets (33%) belonging to MSL. The test data samples are shown in table 1:
TABLE 1
By reporting relevant data records of the SMAP and the MSL through the ISA, the time range and the abnormal type of each set of telemetric data are marked, and as shown in figure 7, the shaded part is the abnormal value of the set of telemetric data.
The detailed total amount of training data, total amount of abnormal sequences, total amount of testing telemetry channels and total amount of testing data are shown in table 2 below:
TABLE 2
LSTM model parameter setting and performance index:
the LSTM parameters used by each training set are the same, the predicted values output by the LSTM model A and the LSTM model B are fused by a weight matrix, and the relevant parameters and fusion parameters of the LSTM model are shown in the following table 3:
TABLE 3
The performance analysis of the anomaly detection method provided by the invention mainly comprises four aspects: accuracy and false detection rate for SMAP point anomaly, accuracy and false detection rate for MSL point anomaly, accuracy and false detection rate for SMAP context anomaly, and accuracy and false detection rate for MSL point anomaly.
When the detected abnormality is overlapped with the marked abnormality in a large range, the abnormality is considered to be successfully detected, and the ratio of the total number of successfully detected abnormal sequences to the total number of actual abnormal sequences is the accuracy of the abnormality detection scheme, as shown in fig. 8a) and 8 b).
When the detected abnormality does not belong to the actual abnormality or the overlap portion is extremely small, it is regarded as a false detection. The ratio of the total number of the sequences detected by mistake to the sum of the number of the sequences detected by mistake and the number of the sequences detected accurately is the false detection rate of the abnormal detection scheme.
As shown in table 4, compared to the LSTM-based anomaly detection method, the integrated LSTM-based telemetry data anomaly detection method has a certain degree of improvement in accuracy in both point anomaly and context anomaly. The improvement of the abnormality detection accuracy rate for the SMAP is particularly obvious. The accuracy of the telemetry data detection method based on the integrated LSTM improves the accuracy of point anomaly detection and context anomaly detection by more than 11%. And aiming at the abnormal detection of the MSL, the accuracy rate of the abnormal context is improved by more than 11%. In a whole view, the method provided by the invention has more obvious improvement on the accuracy rate of the context abnormity, and the detection accuracy rate of the point abnormity is improved by a small extent.
TABLE 4
The above-described calculation examples of the present invention are merely to explain the calculation model and the calculation flow of the present invention in detail, and are not intended to limit the embodiments of the present invention. It will be apparent to those skilled in the art that other variations and modifications of the present invention can be made based on the above description, and it is not intended to be exhaustive or to limit the invention to the precise form disclosed, and all such modifications and variations are possible and contemplated as falling within the scope of the invention.
Claims (5)
1. A spacecraft telemetry data anomaly detection method based on integrated LSTM is characterized by comprising the following steps:
step one, selecting training dataWherein:is a group of m-dimensional vectors represented by a remote control command matrix at the time t and a remote measuring value at the time t, and the remote control command matrix at the time t istelemetry value at time t ofPreprocessing the selected training data S to obtain reconstructed input X;
the specific process of preprocessing the selected training data S to obtain the reconstructed input X is as follows:
splitting the training data S into a plurality of sub-matrices comprising successive time vectors, said sub-matrices being represented respectively as:and the number of columns of each secondary matrix is ltThen the reconstructed input X is represented as:
wherein:the first secondary matrix is represented as a function of,a second sub-matrix is represented which is,representing the nth secondary matrix;
step two, respectively extracting a remote control instruction matrix in each secondary matrix contained in the reconstructed input X, wherein the remote control instruction matrix in each secondary matrix is respectively expressed asRespectively calculating the L2 norm of the remote control command matrix corresponding to each secondary matrix, taking the median of all the calculated L2 norms as a threshold beta, and taking the remote control command with the L2 norm larger than the threshold betaTaking a set formed by secondary matrixes corresponding to the matrixes as a training set A, and taking a set formed by the secondary matrixes corresponding to the remote control instruction matrixes with the L2 norm being less than or equal to a threshold value beta as a training set B;
step three, constructing an integrated LSTM model containing an LSTM model A and an LSTM model B, and respectively training the LSTM model A and the LSTM model B by utilizing the training set A and the training set B obtained in the step two to obtain a well-trained LSTM model A and an LSTM model B;
the LSTM model A in the third step comprises an input layer, a first hidden layer, a second hidden layer and an output layer, the loss function of the LSTM model A adopts an absolute value error function, the optimization function adopts an Adam function, and the value of dropout is [0.2, 0.3 ];
the structure and parameter setting of the LSTM model B are the same as those of the LSTM model A;
the specific form of the absolute value error function is as follows:
wherein: k is the total number of predicted values, yiFor the ith real value, the value of the real value,is the (i) th predicted value,is the absolute value error function value;
step four, test data are preprocessed to obtain a test set, the test set is respectively input into the well-trained LSTM model A and the well-trained LSTM model B, and a telemetering predicted value sequence formed by predicted values of the telemetering values corresponding to each secondary matrix output by the LSTM model A is respectively obtainedTelemetering predicted value sequence formed by predicted values of telemetering values corresponding to each secondary matrix output by LSTM model B
Setting fusion parameters, and determining a final telemetering value prediction sequence according to a telemetering prediction value sequence consisting of prediction values of telemetering values corresponding to each secondary matrix output by the LSTM model A, a telemetering prediction value sequence consisting of prediction values of telemetering values corresponding to each secondary matrix output by the LSTM model B and the fusion parameters
Step five, the final telemetering value prediction sequence determined in the step fourThe value in (d) is subtracted from the actual remote value at the corresponding time to obtain an error sequence e ═ e { (e)(1),e(2),…,e(t),…},e(t)Represents the error at time t;
smoothing the error sequence by adopting an exponential weighted average method to obtain a smoothed error sequence
Grouping the smooth error values in the smooth error sequence, and respectively calculating the dynamic threshold of each groupExceeding dynamic threshold value in each groupAnd the value of the smoothing error is less than the dynamic threshold value in the groupAnd is closest to the dynamic thresholdIs smoothed byThe error value is subtracted, if the difference value is larger than the discrimination threshold value p, the exceeding dynamic threshold value is determinedThe telemetering value at the moment corresponding to the smoothing error value is regarded as an abnormal value; if the difference is less than or equal to the discrimination threshold p, the exceeding dynamic threshold is determinedThe telemetering value of the moment corresponding to the smoothing error value is taken as a normal value;
and step six, repeating the process from the step four to the step five for the telemetering data of the spacecraft to be detected, and realizing the detection of the abnormity of the telemetering data of the spacecraft to be detected.
2. The method for detecting spacecraft telemetry data anomaly based on integrated LSTM according to claim 1, wherein the specific process of the second step is as follows:
respectively extracting a remote control instruction matrix in each secondary matrix contained in the reconstructed input X, and respectively representing the remote control instruction matrix in each secondary matrix asWherein:is composed ofThe matrix of remote control commands in (1),is composed ofThe matrix of remote control commands in (1),is composed ofThe remote control instruction matrix in (1);
wherein:representsThe transpose of (a) is performed,representsThe L2 norm;representing and calculating matrixMaximum eigenvalue of;
After the L2 norm of each remote control command matrix is calculated, the median of all L2 norms is taken as a threshold value beta, a set formed by secondary matrixes corresponding to the remote control command matrix with the L2 norm being less than or equal to the threshold value beta is taken as a training set B, and a set formed by the secondary matrixes corresponding to the remote control command matrix with the L2 norm being greater than the threshold value beta is taken as a training set A.
3. The method for detecting the anomaly of the telemetry data of the LSTM-integrated spacecraft as claimed in claim 1, wherein the training set A and the training set B obtained in the second step are used in the third step to train the LSTM model A and the LSTM model B respectively to obtain a trained LSTM model A and a trained LSTM model B; the specific process comprises the following steps:
respectively inputting the training set A and the training set B into an LSTM model A and an LSTM model B for training, setting the maximum iteration times of the LSTM model A and the LSTM model B for training to be alpha, continuously adjusting each weight matrix of the LSTM model A and the LSTM model B in each iteration training process, and stopping training until the maximum iteration times are reached to obtain the well-trained LSTM model A and the LSTM model B.
4. The method according to claim 1, wherein the method for detecting the anomaly of telemetry data of the LSTM-integrated spacecraft is characterized in that a final telemetry prediction sequence is determined according to a telemetry prediction value sequence formed by prediction values of telemetry values corresponding to each secondary matrix output by the LSTM model A, a telemetry prediction value sequence formed by prediction values of telemetry values corresponding to each secondary matrix output by the LSTM model B and fusion parametersThe method comprises the following steps:
for secondary matrix in test setSecondary matrixThe corresponding remote control command matrix isSecondary matrix of the imageInputting the LSTM model A to obtain the predicted value of the remote measurement value output by the LSTM model ASecondary matrix of the imageInputting the LSTM model B to obtain the predicted value of the remote measurement value output by the LSTM model B
Similarly, the predicted value of the remote measurement value output by the LSTM model A and the predicted value of the remote measurement value output by the LSTM model B are input to other secondary matrixes in the test set;
then the remote measuring predicted value sequence formed by the predicted values of the remote measuring values corresponding to each secondary matrix output by the LSTM model AExpressed as:
then the remote measuring predicted value sequence formed by the predicted values of the remote measuring values corresponding to each secondary matrix output by the LSTM model BExpressed as:
if the remote control instruction matrixL2 norm is greater than threshold β, then secondary matrixCorresponding final telemetry value prediction valuesExpressed as:
if the remote control instruction matrixL2 norm is less than or equal to threshold beta, then secondary matrixCorresponding final telemetry value prediction valuesExpressed as:
5. The method for detecting spacecraft telemetry data anomaly based on integrated LSTM according to claim 1, wherein smoothing is performed on an error sequence by using an exponential weighted average method to obtain a smoothed error sequence es(ii) a The specific process comprises the following steps:
for the smoothed error value at time t,is the smoothing error value at the time t-1, and gamma is the threshold value of the smoothing error;
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910194332.0A CN109934337B (en) | 2019-03-14 | 2019-03-14 | Spacecraft telemetry data anomaly detection method based on integrated LSTM |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910194332.0A CN109934337B (en) | 2019-03-14 | 2019-03-14 | Spacecraft telemetry data anomaly detection method based on integrated LSTM |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109934337A CN109934337A (en) | 2019-06-25 |
CN109934337B true CN109934337B (en) | 2020-12-25 |
Family
ID=66987056
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910194332.0A Active CN109934337B (en) | 2019-03-14 | 2019-03-14 | Spacecraft telemetry data anomaly detection method based on integrated LSTM |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109934337B (en) |
Families Citing this family (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110738311A (en) * | 2019-10-14 | 2020-01-31 | 哈尔滨工业大学 | LSTM network acceleration method based on high-level synthesis |
CN110736968B (en) * | 2019-10-16 | 2021-10-08 | 清华大学 | Radar abnormal state diagnosis method based on deep learning |
CN110968703B (en) * | 2019-12-06 | 2024-01-02 | 国网天津市电力公司电力科学研究院 | Method and system for constructing abnormal metering point knowledge base based on LSTM end-to-end extraction algorithm |
CN111612050B (en) * | 2020-04-30 | 2023-09-15 | 中国西安卫星测控中心 | Method for detecting remote measurement data abnormality |
CN111680355B (en) * | 2020-05-06 | 2022-06-28 | 北京航空航天大学 | Typical telemetering anomaly detection and positioning self-adaptive amplitude geometric mapping method |
CN112257901A (en) * | 2020-09-24 | 2021-01-22 | 北京航天测控技术有限公司 | Abnormity early warning method and device for spacecraft |
CN112287452B (en) * | 2020-10-12 | 2022-05-24 | 哈尔滨工业大学 | Spacecraft maintainability intelligent modeling method |
CN112364304B (en) * | 2020-11-20 | 2023-04-07 | 深圳前海微众银行股份有限公司 | Method and device for detecting solar erosion attack of block chain |
CN112381316B (en) * | 2020-11-26 | 2022-11-25 | 华侨大学 | Electromechanical equipment health state prediction method based on hybrid neural network model |
CN112629905A (en) * | 2020-12-22 | 2021-04-09 | 杭州未名信科科技有限公司 | Equipment anomaly detection method and system based on deep learning and computer medium |
CN114997253A (en) * | 2021-02-23 | 2022-09-02 | 哈尔滨工业大学 | Intelligent state anomaly detection method, monitoring system and monitoring method for satellite constellation |
CN113077110A (en) * | 2021-04-21 | 2021-07-06 | 国家海洋信息中心 | GRU-based harmonic residual segmented tide level prediction method |
CN113162811B (en) * | 2021-06-01 | 2021-12-28 | 长扬科技(北京)有限公司 | Industrial control network flow abnormity detection method and device based on deep learning |
CN113554229A (en) * | 2021-07-23 | 2021-10-26 | 国网青海省电力公司信息通信公司 | Three-phase voltage unbalance abnormality detection method and device |
CN113569424B (en) * | 2021-08-05 | 2022-06-07 | 北京微纳星空科技有限公司 | Satellite operation simulation method and device, electronic equipment and storage medium |
CN113673787B (en) * | 2021-09-10 | 2023-09-26 | 中国舰船研究设计中心 | Unmanned cluster multi-domain detection data track association and prediction method |
CN114254695B (en) * | 2021-11-18 | 2023-04-07 | 中国空间技术研究院 | Spacecraft telemetry data self-adaptive anomaly detection method and device |
CN115099371B (en) * | 2022-08-24 | 2022-11-29 | 广东粤港澳大湾区硬科技创新研究院 | LSTM anomaly detection method, device, equipment and storage medium |
CN115824261B (en) * | 2023-01-15 | 2023-05-23 | 北京理工大学 | Control moment gyroscope fault detection method and device and related storage medium |
CN116680647A (en) * | 2023-08-04 | 2023-09-01 | 中科星图测控技术股份有限公司 | Method for detecting satellite data by constructing LSTM network by using preprocessed data |
CN118133435A (en) * | 2024-05-08 | 2024-06-04 | 北京理工大学长三角研究院(嘉兴) | SVR and clustering-based complex spacecraft on-orbit anomaly detection method |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103679378B (en) * | 2013-12-20 | 2016-08-31 | 北京航天测控技术有限公司 | Method and device based on telemetry assessment heath state of spacecraft |
CN103974311B (en) * | 2014-05-21 | 2017-06-20 | 哈尔滨工业大学 | Based on the Condition Monitoring Data throat floater detection method for improving Gaussian process regression model |
CN107480182B (en) * | 2017-07-05 | 2020-01-03 | 中国科学院光电研究院 | Spacecraft telemetry data tracing method and system |
CN108388795A (en) * | 2018-02-11 | 2018-08-10 | 浙江工业大学 | A kind of confrontation attack defense method based on LSTM detectors |
CN108960303B (en) * | 2018-06-20 | 2021-05-07 | 哈尔滨工业大学 | Unmanned aerial vehicle flight data anomaly detection method based on LSTM |
-
2019
- 2019-03-14 CN CN201910194332.0A patent/CN109934337B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN109934337A (en) | 2019-06-25 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109934337B (en) | Spacecraft telemetry data anomaly detection method based on integrated LSTM | |
CN113777496B (en) | Lithium ion battery residual life prediction method based on time convolution neural network | |
CN116757534B (en) | Intelligent refrigerator reliability analysis method based on neural training network | |
Lindemann et al. | Anomaly detection and prediction in discrete manufacturing based on cooperative LSTM networks | |
CN113298297B (en) | Wind power output power prediction method based on isolated forest and WGAN network | |
CN109471698B (en) | System and method for detecting abnormal behavior of virtual machine in cloud environment | |
CN113988210A (en) | Method and device for restoring distorted data of structure monitoring sensor network and storage medium | |
CN114397043B (en) | Multi-point temperature intelligent detection system | |
CN115271186B (en) | Reservoir water level prediction and early warning method based on delay factor and PSO RNN Attention model | |
CN114254695B (en) | Spacecraft telemetry data self-adaptive anomaly detection method and device | |
CN114548509A (en) | Multi-type load joint prediction method and system for multi-energy system | |
CN112381282A (en) | Photovoltaic power generation power prediction method based on width learning system | |
CN110991776A (en) | Method and system for realizing water level prediction based on GRU network | |
CN116431966A (en) | Reactor core temperature anomaly detection method of incremental characteristic decoupling self-encoder | |
CN114091600B (en) | Data-driven satellite association fault propagation path identification method and system | |
CN110222825B (en) | Cement product specific surface area prediction method and system | |
CN113688770B (en) | Method and device for supplementing long-term wind pressure missing data of high-rise building | |
CN111754033B (en) | Non-stationary time sequence data prediction method based on cyclic neural network | |
US20230350402A1 (en) | Multi-task learning based rul predication method under sensor fault condition | |
CN116415505A (en) | System fault diagnosis and state prediction method based on SBR-DBN model | |
CN111177916A (en) | Satellite electromechanical product service life prediction method based on machine deep learning | |
CN114399024B (en) | Oil gas concentration big data intelligent detection system | |
CN114358244B (en) | Big data intelligent detection system of pressure based on thing networking | |
CN115204362A (en) | Method for predicting residual life interval of mechanical equipment | |
CN113569324A (en) | Slope deformation monitoring abnormal data analysis and optimization method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |