CN113391622B

CN113391622B - Spacecraft attitude system anomaly detection method using multivariate multistep prediction technology

Info

Publication number: CN113391622B
Application number: CN202110702313.1A
Authority: CN
Inventors: 刘亚杰; 沈凯丽; 张涛; 余京; 黄生俊; 雷洪涛; 王锐; 史志超
Original assignee: National University of Defense Technology
Current assignee: National University of Defense Technology
Priority date: 2021-06-22
Filing date: 2021-06-22
Publication date: 2022-09-02
Anticipated expiration: 2041-06-22
Also published as: CN113391622A

Abstract

The method comprises the steps of improving a long-term and short-term memory model by using an attention mechanism to obtain a multi-element multi-step prediction neural network, predicting data obtained at multiple historical moments of different parameters of a plurality of spacecraft attitude systems by the neural network to obtain multiple predicted data at next moments, calculating according to actual data at the next moments and predicted data at corresponding moments to obtain errors between the actual data and the predicted data, and comparing the errors with error thresholds of corresponding parameters to realize abnormal detection of the spacecraft attitude systems. By adopting the method, the real-time monitoring of the telemetering parameters of the attitude system can be effectively realized, so that the abnormity detection is carried out.

Description

Spacecraft attitude system anomaly detection method using multivariate multistep prediction technology

Technical Field

The application relates to the technical field of anomaly detection of spacecraft telemetry data, in particular to a spacecraft attitude system anomaly detection method utilizing a multi-element multi-step prediction technology.

Background

The development of aerospace technology has driven the expansion of the human activity field from the atmosphere to the universe. The spacecraft is a large multifunctional complex system which comprehensively utilizes the advanced knowledge in the multidisciplinary fields of electronics, machinery, materials, energy, control science, computer science and the like, and the development of the spacecraft needs to consume a large amount of manpower and resources. With the rapid development of information technology, the functions of large complex systems such as spacecraft are continuously improved and increased, and the components are more and more complex, the combination and cooperation among the components are more difficult, and system abnormity and even failure can be caused once problems occur. In addition, in the complex and unknown space of the spacecraft, various kinds of abnormality may occur to the system due to the influence of various unavoidable factors which are difficult to predict, such as radiation, space debris and the like, so that the reliability of the system is influenced, the system is difficult to complete the predetermined function, and even serious economic loss and catastrophic accidents are caused.

The satellite is the spacecraft with the largest number and bears important scientific research and practical application tasks such as scientific detection, communication, meteorological observation, navigation, reconnaissance and the like. The satellite is a complex controlled object, is used as an attitude control system of a most important control subsystem of the satellite, and mainly bears two tasks of attitude determination and attitude control, namely acquiring the position of the satellite in a state, controlling the satellite to operate according to a preset orbit and an attitude and ensuring the satellite to complete flight and work tasks. According to statistics, compared with other control systems, the attitude control system is more susceptible to the influence of outer space to cause failure, directly influences the operation of a satellite, and is more serious in harm. Therefore, an attitude system abnormity detection model is constructed, the operating state of the attitude system is monitored in real time, warning is given when abnormity occurs, manual investigation is carried out, major faults are avoided, and normal operation of the attitude system is guaranteed.

The attitude system anomaly detection mainly comprises two aspects, namely firstly detecting whether an anomaly occurs, and then positioning the cause and the part of the anomaly. In practical applications, threshold value detection based on signal monitoring, consistency and correlation detection of redundant accessories or functional related components are still widely used, but the former may be affected by noise and environmental changes to cause false detection, and the latter redundancy requires additional equipment and corresponding maintenance cost, so that anomaly detection methods based on models, data and the like are valued by experts of scholars. The current leading research is mainly divided into three major categories: model-based, data-based, hybrid model-based, and data-based anomaly detection.

Most control systems are based on physical and mathematical modeling, so that the anomaly detection based on the model has scientific theoretical basis and is applied to satellite attitude systems in recent decades. The method comprises the steps of constructing a mathematical model of a system based on the abnormal detection of the model, generating residual errors by using methods such as state estimation, parameter estimation, equivalent space equations and the like, and then analyzing and carrying out the abnormal detection based on the residual errors. However, as satellite control systems become more complex, it becomes increasingly difficult for conventional modeling methods to accurately model components or the whole. And the uncertainty of the environment causes that the modeling-based anomaly detection method has weak adaptability, and is increasingly difficult to deal with multivariable and nonlinear complex systems such as attitude systems. On one hand, research is biased towards more complex modeling such as state/parameter synchronous estimation and integration based on multiple models, robust modeling considering uncertainty factors, and the like, and on the other hand, data-driven modeling and anomaly detection are also attracting attention in the industry.

Data-driven anomaly monitoring is successfully applied to industrially complex systems. For gesture systems, data-driven exceptions are increasingly gaining attention from industry and academia. For the arrangement of the existing research at present, unit abnormality detection and multiple abnormality detection can be carried out in large quantity according to the type of abnormality, namely single-parameter abnormality or multi-parameter abnormality. Unit anomalies are typically analyzed for a certain parameter or class of parameters, and gesture systems have less research on this class. The posture system has numerous parameters and complex modes, so that the multivariate abnormality of the posture system is researched more. The multivariate anomaly detection method can be divided into two categories of supervised learning and unsupervised learning. And if the abnormal data are labeled in advance, performing abnormal recognition on the abnormal data by using a classifier such as machine learning. For unlabeled anomalies, with the development of machine learning and deep learning, anomaly detection based on data driving is also recognized and applied more and more by virtue of excellent detection effect and adaptability to uncertainty.

The hybrid model is the combination of the former two methods, and combines a data-driven method model with an observer filter and the like, so that the scientific rigor of the model and the high efficiency of the data are both considered. In the case of obtaining an accurate mathematical model, the observation-based method is directly effective, but in practical use, the influence on model uncertainty, time lag, modeling error, and the like is not negligible. Compared with the traditional state estimation method which needs accurate modeling based on a physical system, the method has the advantages that the state is reconstructed, and the strong nonlinear fitting capability of the neural network is utilized for a lot of research, so that the neural network is fused with accurate modeling to construct the observer. Fusion models are currently less studied due to their cross-domain difficulty.

Disclosure of Invention

Therefore, in order to solve the technical problems, it is necessary to provide a spacecraft attitude system anomaly detection method using a multi-element multi-step prediction technology, which aims at the characteristics of a satellite attitude control system, such as complex structure, numerous parameters, many working state modes, complex features and the like.

A spacecraft attitude system anomaly detection method utilizing a multi-element multi-step prediction technology comprises the following steps:

acquiring sample data of a plurality of parameters of a spacecraft attitude system, wherein the sample data is a plurality of data which are continuously arranged in a time sequence within a period of time;

performing stationarity analysis and correlation analysis on each sample data, selecting a plurality of steady state parameters and auxiliary parameters related to each steady state parameter, and constructing a training data set according to the sample data corresponding to the steady state parameters and the auxiliary parameters;

carrying out standardization processing on the training data set to obtain a standard data set;

constructing a multi-element multi-step prediction neural network based on an attention mechanism and a long-short term memory neural network, inputting the standard data set into the multi-element multi-step prediction neural network for training to obtain a trained multi-element multi-step prediction neural network, wherein the trained multi-element multi-step prediction neural network has the capability of predicting data of a time period after a parameter according to data of a time period before the parameter;

calculating to obtain an error threshold of each steady state parameter by adopting a 3 sigma rule according to the predicted data and the real data of each steady state parameter;

acquiring actual data of historical time of a plurality of parameters to be detected in a spacecraft attitude system, and carrying out standardization processing on the actual data to obtain standardized actual data;

inputting the standardized actual data into a trained multivariate multi-step prediction neural network to obtain real-time prediction data of each parameter to be detected in the next time period;

and acquiring real-time real data of each parameter to be detected, calculating errors between the real-time real data of each parameter to be detected and real-time prediction data at a corresponding moment respectively, and comparing the errors with error thresholds of corresponding parameters to detect whether the current spacecraft attitude system is abnormal or not.

In one embodiment, the selecting the plurality of stationary state parameters and the auxiliary parameter related to each stationary state parameter by performing stationary analysis and correlation analysis on each sample data includes:

performing stationarity analysis on each sample data, and selecting a parameter with a steady state from the plurality of parameters as the plurality of steady state parameters;

and performing correlation analysis according to the sample data and the steady state parameters, and selecting parameters related to the steady state parameters from the parameters as the auxiliary parameters.

In one embodiment, the inputting of the training data set into the multivariate multi-step predictive neural network for training comprises:

the long-term and short-term memory neural network comprises a plurality of processing layers which are connected in sequence, and each processing layer receives sample data at different moments as input data and outputs output data corresponding to the processing layer after corresponding processing;

obtaining vector weights output by the processing layer at each time step based on attention mechanism matching according to the output data of the processing layer time step and the corresponding real data;

and after carrying out weighted calculation on the vector weight output by each time step of the processing layer and the output data of each time step of the processing layer, integrating the vector weight and the output data of each time step into the final prediction calculation of the multi-element multi-step prediction neural network.

In one embodiment, before the training of the multivariate multistage predictive neural network, the method further includes:

presetting the time series length of the sample data input by the multi-element multi-step prediction neural network, and presetting the time series length of the preset data output by the multi-element multi-step prediction neural network.

In one embodiment, before inputting the normalized actual data into the trained multivariate multi-step predictive neural network, the method further comprises:

presetting the time series length of the actual data input by the multi-element multi-step prediction neural network, and presetting the time series length of the actual preset data output by the multi-element multi-step prediction neural network.

In one embodiment, the calculating the error threshold of each steady-state parameter according to the predicted data and the actual data of each steady-state parameter by using 3 sigma rule comprises:

calculating an error time sequence of each steady state parameter within a period of time according to each steady state parameter prediction data and real data;

and calculating an error threshold value of each steady state parameter by adopting a 3 sigma rule according to the mean value and the variance of each error time sequence.

An apparatus for anomaly detection of a spacecraft attitude system using a multivariate multistage prediction technique, the apparatus comprising:

the system comprises a sample data acquisition module, a data processing module and a data processing module, wherein the sample data acquisition module is used for acquiring sample data of a plurality of parameters of the spacecraft attitude system, and the sample data is a plurality of data which are continuously arranged in a time period by taking time as a sequence;

the training data set construction module is used for performing stationarity analysis and relevance analysis on each sample data, selecting a plurality of steady state parameters and auxiliary parameters related to each steady state parameter, and constructing a training data set according to the sample data corresponding to the steady state parameters and the auxiliary parameters;

the standardization processing module is used for carrying out standardization processing on the training data set to obtain a standard data set;

the neural network training module is used for constructing a multi-element multi-step prediction neural network based on an attention mechanism and a long-short term memory neural network, inputting the standard data set into the multi-element multi-step prediction neural network for training to obtain a trained multi-element multi-step prediction neural network, and the trained multi-element multi-step prediction neural network has the capability of predicting data of a time period after a parameter according to data of a time period before the parameter;

the error threshold calculation module is used for calculating the error threshold of each steady state parameter according to the prediction data and the real data of each steady state parameter by adopting a 3 sigma rule;

the actual data acquisition module is used for acquiring actual data of historical time of a plurality of parameters to be detected in the spacecraft attitude system and carrying out standardization processing on the actual data to obtain standardized actual data;

the actual prediction data obtaining module is used for inputting the standardized actual data into a trained multi-element multi-step prediction neural network to obtain real-time prediction data of each parameter to be detected in the next time period;

and the system anomaly detection module is used for acquiring real-time real data of each parameter to be detected, respectively calculating errors between the real-time real data of each parameter to be detected and real-time predicted data at a corresponding moment, and then comparing the errors with error thresholds of corresponding parameters to detect whether the current spacecraft attitude system is anomalous.

A computer device comprising a memory and a processor, the memory storing a computer program, the processor implementing the following steps when executing the computer program:

acquiring sample data of a plurality of parameters of a spacecraft attitude system, wherein the sample data is a plurality of data which are continuously arranged in a time sequence within a period of the parameters;

and acquiring real-time real data of each parameter to be detected, respectively calculating an error between the real-time data of each parameter to be detected and real-time predicted data at a corresponding moment, and comparing the error with an error threshold of the corresponding parameter to detect whether the current spacecraft attitude system is abnormal.

A computer-readable storage medium, on which a computer program is stored which, when executed by a processor, carries out the steps of:

and acquiring real-time real data of each parameter to be detected, calculating an error between the real-time real data of each parameter to be detected and actual real-time predicted data at a corresponding moment, and comparing the error with an error threshold of the corresponding parameter to detect whether the current spacecraft attitude system is abnormal.

According to the spacecraft attitude system anomaly detection method using the multivariate multistep prediction technology, the long-term and short-term memory model is improved by using an attention mechanism to obtain the multivariate multistep prediction neural network, the neural network can predict data obtained at a plurality of historical moments of different parameters of a plurality of spacecraft attitude systems to obtain predicted data at the next moments, so that an error between the actual data at the next moment and the predicted data at the corresponding moments is obtained by calculation, and the error is compared with an error threshold value of the corresponding parameter to realize anomaly detection of the spacecraft attitude system. The method can effectively realize the real-time monitoring of the telemetering parameters of the attitude system, thereby carrying out the anomaly detection.

Drawings

FIG. 1 is a schematic flow chart of a method for anomaly detection of a spacecraft attitude system in one embodiment;

FIG. 2 is a diagram of the processing of data in an LSTM in one embodiment;

FIG. 3 is a diagram illustrating an Attention mechanism in one embodiment;

FIG. 4 is a schematic diagram of a multivariate multistage predictive neural network in one embodiment;

FIG. 5 is a schematic diagram of an exemplary multi-step time-series prediction input-output array;

FIG. 6 is a schematic diagram of the steps of a spacecraft attitude system anomaly detection method in a data experiment;

FIG. 7 is a visualization of periodic parameters of the attitude system in step one of the data experiments;

FIG. 8 is a thermodynamic diagram of the correlation of pose system parameters in step one of the data experiments;

FIG. 9 is a graph showing the prediction error for each of the four steps in the data experiment;

FIG. 10 is a schematic diagram of three types of anomalies that were manually constructed in step six of the data experiment;

FIG. 11 is a schematic diagram of the identification of constant value anomalies and time varying anomalies by the four models in step eight of the data experiment;

FIG. 12 is a schematic diagram illustrating identification of abnormal deadlock by four models in step eight of the data experiment;

FIG. 13 is a block diagram of an anomaly detection apparatus for a spacecraft attitude system in accordance with an embodiment;

FIG. 14 is a diagram illustrating an internal structure of a computer device according to an embodiment.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.

As shown in fig. 1, the application provides a spacecraft attitude system anomaly detection method using a multivariate multi-step prediction technology, which includes the following steps:

s100, obtaining sample data of a plurality of parameters of the spacecraft attitude system, wherein the sample data is a plurality of data which are continuously arranged in a time sequence within a period of time by taking time as a parameter;

step S110, performing stationarity analysis and relevance analysis on each sample data, selecting a plurality of steady state parameters and auxiliary parameters related to each steady state parameter, and constructing a training data set according to sample data corresponding to the steady state parameters and the auxiliary parameters;

step S120, carrying out standardization processing on the training data set to obtain a standard data set;

step S130, constructing a multi-element multi-step prediction neural network based on an attention mechanism and a long-short term memory neural network, inputting a training data set into the multi-element multi-step prediction neural network for training to obtain a trained multi-element multi-step prediction neural network, wherein the trained multi-element multi-step prediction neural network has the capability of predicting data of a time period after a parameter according to data of a time period before the parameter;

step S140, calculating to obtain an error threshold of each steady state parameter according to the prediction data and the real data of each steady state parameter by adopting a 3 sigma rule;

s150, acquiring actual data of historical time of a plurality of parameters to be detected in the spacecraft attitude system, and carrying out standardization processing on the actual data to obtain standardized actual data;

step S160, inputting the standardized actual data into the trained multivariate multi-step prediction neural network to obtain the real-time prediction data of each parameter to be detected in the next time period;

step S170, acquiring real-time real data of each parameter to be detected, calculating an error between the real-time real data of each parameter to be detected and real-time predicted data of a corresponding moment, and comparing the error with an error threshold of the corresponding parameter to detect whether the current spacecraft attitude system is abnormal.

In this embodiment, steps S100 to S140 train a multi-element multi-step prediction neural network constructed based on an attention mechanism and a long-and-short-term memory neural network using a training data set, and calculate an error threshold of each steady-state parameter, and steps S150 to S170 predict data of multiple parameters at multiple times of a next period using the trained multi-element multi-step prediction neural network, and determine whether a current attitude system is abnormal or not by using an error between the predicted data and actually obtained data.

In this embodiment, the multiple steps refer to predicting data at a plurality of next time instants by using data received at a plurality of historical time instants by using a plurality of parameters.

In step S100, the acquired sample data is historical telemetry data of different parameters in the spacecraft attitude system, and the data in the sample data is normal data. Each data sample includes data received at each time within a period of time, i.e., time series data.

In step S110, performing stationarity analysis and correlation analysis on each sample data to select a plurality of stationary state parameters and auxiliary parameters related to each stationary state parameter includes: and performing stationarity analysis on each sample data, selecting a parameter with a steady state from the plurality of parameters as a plurality of steady state parameters, performing correlation analysis according to each sample data and the plurality of steady state parameters, and selecting a parameter related to each steady state parameter from the plurality of parameters as an auxiliary parameter.

Specifically, firstly, stationarity analysis is performed on each sample data, and stationarity is one of basic tests for establishing a time series prediction model. Time-series predictions are based on history and predictions made in the future, so the data needs to have some basic characteristics to remain unchanged, i.e. the data is smooth, before predictions are meaningful.

The stability analysis method includes but is not limited to: ADF root mean square test, rolling statistics, etc. The ADF unit root test has unit roots in the time sequence, the obtained t statistic is compared with the statistic value of rejecting the original hypothesis in different degrees (1%, 5%, 10%), if the t statistic is less than 1%, the original hypothesis can be rejected on a 99% confidence interval, and the time sequence is considered to be stable; and when the t statistic is between 1% and 5% rejection of the original hypothesis, rejecting the original hypothesis at a 95% confidence interval, and determining whether the time sequence is stable or not according to the test requirement. Whereas the parameter p-value requires less than a given significance level, preferably very close to 0, roll statistics, etc.

The parameters with stable states in the parameters can be selected through stability analysis, and because the parameters in the space-time attitude system are mutually influenced, and the parameters related to the stable-state parameters are selected from the rest parameters through correlation analysis to assist in prediction, the multi-element multi-step prediction neural network obtained subsequently can be predicted more accurately.

Specifically, the Correlation analysis mainly uses Pearson Correlation coefficient (also commonly referred to as r value) for measuring the degree of Correlation between variables (i.e., between parameters). After the data pass the significance test, the higher the correlation coefficient is, the more closely the relationship between the two is. The Pearson correlation coefficient formula is as follows:

in the formula (1), X _i And Y _i Representing the corresponding point location values in both variables. The value range of r is between 1 and-1, the more the value is close to 1, the more positive correlation is, and the more negative correlation is, the more close to-1; when r is close to 0, the lower the correlation degree.

In step S120, a training data set constructed by the selected plurality of stationary state parameters and the sample data corresponding to the relevant auxiliary parameters is normalized to obtain a standard data set with a mean value of 0 and a variance of 1.

Specifically, the data normalization process is as follows:

suppose Z is [ Z ] ₁ ,z ₂ ,...,z _N ] ^T ∈R ^m Representing a set of nxm high dimensional data, wherein m represents the number of parameters, each parameter comprising s independent sample samples, constructing a telemetry parameter data matrix, wherein each column represents a parameter variable and each row represents a sample data. Carrying out standardization processing on the matrix Z to obtain a matrix X ═ X _ij ]Wherein:

the mean value and standard deviation calculation mode of each parameter are respectively as follows:

after the data is subjected to the standardization processing, the method also comprises the following steps before the multivariate multi-step prediction neural network is trained: presetting the time sequence length of the sample data input by the multi-element multi-step prediction neural network, and presetting the time sequence length of the preset data output by the multi-element multi-step prediction neural network, namely, setting the input and output step length of the multi-element multi-step prediction neural network. For example, the sample data of each parameter includes data corresponding to 20 consecutive time points, and if the input step size of the neural network is set to 10 and the output step size is set to 5, the data corresponding to the previous 10 time points is input into the neural network and the predicted value for the next 5 time point data is output.

In the subsequent training of the multi-element multi-step prediction neural network, in order to improve the accuracy of the neural network, the standard data set may be divided into a divided training set and a test set, in one embodiment, the standard data in the training set accounts for 80% of the standard data set, and the standard data in the test set accounts for 20% of the standard data set. Therefore, the multivariate multi-step prediction neural network is trained by utilizing the training set to obtain a preliminarily finished neural network, and then the prediction capability of the neural network is adjusted by utilizing the test set so that the prediction capability of the multivariate multi-step prediction neural network reaches an expected value.

In step S130, the multi-step predictive neural network is actually a neural network obtained by improving the long-short term memory neural network based on an attention mechanism.

Wherein, the long-short term memory neural network is also called long-short term memory model, which is called LSTM model for short. The main idea of the model is to introduce an adaptive gating mechanism, determine how much the LSTM unit remains in the previous state, and remember the extracted features of the current data input.

During the training process, each LSTM unit, as shown in fig. 2, learns how to weigh its input elements (input gates), adjust its contribution to memory (input modulator), remove the memory cell weights (forget gate), and release the memory weights (output gate). The data processing procedure in each LSTM unit is as follows:

f _t ＝σ(W _f ·[h _t-1 ,x _t ]+b _f ) (2)

i _t ＝σ(W _i ·[h _t-1 ,x _t ]+b _i ) (4)

o _t ＝σ(W _o [h _t-1 ,x _t ]+b _o ) (6)

h _t ＝o _t *tanh(C _t ) (7)

in equation (2-7), t is the new time, x _t For data in the externally entered standard data set, h _t-1 Output for the last moment (short duration memory), C _t-1 For the control of the long-term memory at the previous moment,

new information at the present moment, C _t For new long term memory, h _t For the memory of interest at this moment, σ is the activation function; w _i ，W _c ，W _f ，W _o Respectively the weights of the input gate, the memory gate, the forgetting gate and the output gate, b _i ，b _c ，b _f ，b _o The deviation of the input gate, the memory gate, the forgetting gate and the output gate is respectively; f. of _t Is a forgetting gate, i _t Is an input gate, O _t Is an output gate.

As shown in fig. 3, the Attention mechanism is an Attention mechanism, and the idea of Attention is to determine the weight distribution of the input sequence according to the current state. The formula is as follows, where H ═ H ₁ ,h ₂ ,...,h _t ]Alpha is the weight for the state of the hidden layer. Where w is the parameter vector, t is the step size, H ∈ R ^d×t W, α, R are each R ^d×1 ,R ^1×t ,R ^d×1 The weight is obtained by comparing each element of the sequence with other elements to determine the score of each function, then normalizing, and finally outputting the result of multiplying the original hidden layer by the corresponding weight.

M＝tanh(H) (8)

α＝softmax(w ^T M) (9)

r＝Hα ^T (10)

h ^* ＝tanh(r) (11)

And then, improving an LSTM model based on an Attention mechanism to obtain a multi-step prediction neural network, wherein the improved LSTM model (A-LSTM) is added with the Attention mechanism on the basis of the traditional LSTM model, and the Attention carries out weight distribution calculation on an LSTM hidden layer output sequence to further capture important information and improve the LSTM network prediction effect.

Specifically, in the formula (8), the hidden layer H state is processed by an activation function tanh, and the activation function changes the linearity into the nonlinearity, so that the effect is better. Then, by the formula (9), the weight is obtained. R in equation (10) is the hidden layer and weight multiplication weighting. H in equation (11) is output after reactivation function. The advantage of neural networks over linear fitting is that the fitting ability is strong non-linearly, which is achieved by activating functions.

In step S130, inputting the training data set into the multivariate multi-step predictive neural network for training includes: the long-term and short-term memory neural network comprises a plurality of processing layers (namely LSTM units) which are connected in sequence, wherein each processing layer receives sample data at different moments as input data and outputs output data corresponding to the processing layer after corresponding processing is carried out;

obtaining the vector weight output by each time step of the processing layer based on attention mechanism matching according to the output data of each time step of the processing layer and the corresponding real data;

and after carrying out weighted calculation on the vector weight output by each time step of the processing layer and the output data of each time step of the processing layer, integrating the vector weight and the output data of each time step into the final prediction calculation of the multi-step prediction neural network.

As shown in FIG. 4, the self-loop of the LSTM model is developed, which works by continuously receiving each time step

And output H _t The LSTM then performs the cycle and retention of information in the temporal dimension. The Attention mechanism is applied to H of each time step (i.e., time point) _t Matching with the final true value Y to obtain the matching vector weight alpha of each time step _t And finally, the weights and the time step weights are blended into the calculation of a final prediction sequence, and attention distribution coefficients, namely weight parameter layers, are given based on the predicted values, so that the prediction effect of the LSTM is improved.

Specifically, the LSTM model is a multi-step prediction, each step yields a result, and the result is predicted together with the next step data to obtain the predicted result of the next step. And circulating in this way, and finally obtaining the predicted value as the final result. However, after adding attention, not only the result of the last step but also the results of the previous steps are used, and only the weight is given according to the predicted value. Since the last time step has the greatest impact on the prediction outcome in the LSTM mode, but the actual profile is not necessarily.

According to the method, important parameters needing to be predicted are selected through stability analysis by performing data visualization, feature analysis, correlation analysis and the like on all the parameters. Since the description of the pose system requires a set of parameters instead of one, the prediction thereof is not a single parameter prediction, but a prediction of multivariate parameters. Based on attitude dynamics and attitude kinematics, the attitude information of the satellite is influenced by angular velocity, rotating speed and the like, so that on the basis of predicting the group of parameter time series, relevant parameters of dynamics and kinematics are selected through correlation analysis and finally added into a model to assist in prediction.

And the input and output are constructed by setting the multivariate multi-step prediction neural network input step length and the prediction step length, as shown in fig. 5, for example, three steady state parameters are predicted, 5 auxiliary parameters are added, and the last 10 steps are predicted by the first 20 steps, so that q is 8, k is 3, m is 10, and n is 20. Since the processes of receiving and transmitting data by the satellite, decoding and adding codes and the like also need to consume time, the single-step prediction usually lags behind the generation of actual data and has no practical significance, and therefore, the multi-step prediction is adopted.

In step S140, calculating an error threshold of each steady-state parameter according to the predicted data and the actual data of each steady-state parameter by using 3 sigma rule includes: calculating an error time sequence of each steady state parameter within a period of time according to each steady state parameter prediction data and the real data; and calculating an error threshold value of each steady state parameter by adopting a 3 sigma rule according to the mean value and the variance of each error time sequence.

Specifically, the error threshold for each parameter is calculated as:

∈＝μ(e _s )+zσ(e _s ) (12)

wherein e _s ＝{e ⁰ ,e ¹ ,e ² ,...e ^s The error sequence for each parameter over time,

for the error comparison of the predicted value and the actual value at the current moment, the mean value and the variance of the error time series are respectively calculated as

Based on statistics, a 3 sigma rule exists, when data obey normal distribution, calculation is carried out according to the normal distribution, the probability of the data outside the 3 sigma of the average value is P (| x-mu | > 3 sigma) ≦ 0.0003, the data is an extremely small probability event, and the data cannot occur under normal conditions. Therefore, when the error of the predicted value and the actual value differs from the average error by more than 3 σ, the sample is determined to be an abnormal value. Based on the 3 σ rule, the value of z can be set to 3 in general. According to the actual requirement of the abnormality detection precision, if the parameter requires higher monitoring precision, z can be properly reduced, otherwise, z can be properly increased.

Thus, error thresholds corresponding to the respective steady state parameters can be calculated.

In steps S150 to S17, the anomaly detection is performed on the space attitude system on line in real time.

In step S150, actual data of a group of parameters (multiple parameters) to be detected in a past period of time is received online, each actual data is normalized according to the method in step S120, the normalized actual data is input into the trained multi-element multi-step predictive neural network for prediction, before that, the time series length of the actual data input by the multi-element multi-step predictive neural network is preset, and the time series length of the actual preset data output by the multi-element multi-step predictive neural network is preset. For example, the actual data of the plurality of parameters acquired at present is the time sequence acquired at 20 consecutive time points in the past period of time, and the predicted actual data corresponding to one time point in the future or a plurality of consecutive time points can be obtained by setting the step size of the output.

Thus, in step S170, when the real-time real data of the current time is obtained, the error calculation may be performed on the real-time real data of the current time and the predicted real-time data of the corresponding time, and then the calculated error corresponds to the error threshold of the corresponding parameter, and if the error exceeds the error threshold, an abnormal warning may be performed, so as to implement the real-time detection of the aerospace attitude system.

As shown in fig. 6-12, data experiments based on the above-described anomaly detection of spacecraft attitude systems using multivariate multi-step prediction techniques are presented.

The data set obtained in this experiment is the output current data of resource three satellite from 2016 for 6 months to 2017 for 10 months (data loss of about 1 half and a half in the middle). And processing the data sets to obtain a degradation estimation result of the battery array. As shown in fig. 6, the method specifically includes the following steps:

step 1: selecting parameters;

step 1.1: data visualization is performed on attitude system parameters (magnetometer, star-sensitive quadruple, attitude angle to ground and the like), and partial periodic parameter visualization is shown in fig. 7. The small period is about 98 seconds. The magnetometer (vector type magnetic sensor) is used for measuring the magnitude and direction of the geomagnetic field, namely the component of the geomagnetic field intensity vector in the system where the spacecraft is located) also has a large period, about one day and 1440 seconds.

Step 1.2: data stability analysis is carried out based on ADF unit root inspection, and the data are stable and predictable. Based on the ADF unit root test partial results (see table 1), these four parameters are stable in the long term, magnetometer Y (ZK052089) is not stable in the short term, momentum wheel 1 spin control (ZK052067) is nearly unstable, and star sensor a quad-element 2(ZK052076) and ground attitude angle X (ZK052000) are very stable in both the long and short term.

TABLE 1 ADF Unit root test

Step 1.3: and (5) carrying out correlation analysis. A simple correlation analysis was performed using Pearson coefficients for a total of 112 dimensional parameters of the pose system, the results of which are shown in fig. 8. A closer correlation coefficient to 1 indicates a higher correlation. Some parameters and other parameters can not be subjected to correlation analysis, and data is checked and found, and the parameters are constant in the mode and do not change along with time.

Step 1.4: parameter selection is performed based on the analysis. The experiment is multi-step prediction of a multi-element time sequence, the star sensitive quaternion is a stable time sequence with stable periodicity, and the magnetometer comprises a stable sequence and a non-stable sequence, so that three parameters of the magnetometer X (ZK052088), the magnetometer Y (ZK052088) and the magnetometer Z (ZK052090) are selected for prediction. Based on correlation analysis, five parameters of momentum wheel 1 rotation speed control (ZK052067), momentum wheel 3 rotation speed control (ZK052069), star sensitive A quaternion 1(ZK052075), fine sensitive 1 solar angle 1(ZK052084) and fine sensitive 1 solar angle 2(ZK052085) are selected as correlation parameters to participate in prediction.

Step 2: and performing standardization processing on the sample data corresponding to the selected parameters, wherein the processing method is the same as that in the step S120.

And step 3: establishing a multi-parameter multi-step prediction model based on improved LSTM;

the LSTM network has more hyper-parameters which need to be set in advance, so that when a plurality of LSTMs are subjected to contrast tests, the hyper-parameters are subjected to variable control to ensure the fairness and the training effect of the contrast tests. Specific parameter settings are shown in table 1:

TABLE 1 LSTM hyper-parameter settings

And 4, step 4: and training the standardized training data set based on the multivariate multi-step prediction neural network. With multi-parameter multi-step prediction, the input-output setup is as shown in fig. 5, for example, three steady-state parameters are predicted, 5 auxiliary parameters are added, and the last 10 steps are predicted with the first 20 steps, then q is 8, k is 3, m is 10, and n is 20.

On one hand, sufficient time needs to be reserved for decoding and prediction, and engineering hopes are that the number of steps can be predicted as many as possible; on the other hand, as the number of prediction steps increases, errors accumulate. And therefore a suitable number of prediction steps is required. The data with the step size of the first 20 steps is input, the error of the predicted 20 steps is shown in fig. 9, the error after the predicted 10 steps grows too fast, and the predicted 10 steps can meet the time requirement of decoding and calculation, so the predicted 10 steps are based on the first 20 steps in the experiment. 80% of the data were used as training set and 20% as test set.

The prediction evaluation index adopts average absolute error (MAE), RMSE (root mean square error) and MSE (mean square error) to evaluate the quality of a prediction result. The calculation formula is as follows, wherein

To predict value, y _i As true values:

to test the effectiveness of the method proposed herein, PA-LSTM (plus correlation parameters and attention mechanism) was added with three models for comparison, respectively: P-LSTM (with correlation parameters), A-LSTM (with correlation mechanism but no correlation parameters), LSTM (with no correlation parameters and no correlation mechanism).

The results are shown in Table 2, with PA-LSTM performing best and LSTM alone performing worst. From the perspective of prediction error, the addition of the Attention layer and related parameters can improve the prediction accuracy to a certain extent.

TABLE 2 comparison of predicted Effect of various LSTM models

And 5: error threshold calculation, the mean and variance of the errors of the three parameters obtained by 3 sigma law and the final threshold are shown in table 3:

TABLE 3 mean variance and error threshold of each parameter

Step 6: and (3) carrying out standardization processing on the telemetry data sample to be detected online, and manually adding abnormity for detecting the abnormity detection effect in the same step 1:

1) the abnormal constant deviation means that the actual output and the theoretical output of the signal are different by a constant deviation value within a period of time, and the input deviation or the measurement drift possibly occurs, and the abnormal constant deviation is frequently generated in analog devices. The description form is: y is _out ＝y ₀ +f _c Wherein y is _out To the actual output, y ₀ For ideal output, f _c The deviation is a constant value, as shown in fig. 10 (a).

2) The time-varying anomaly is a time-varying characteristic with a large deviation between an actual output signal and a theoretical value, and may be a regular deviation or an irregular disturbance. The cause may be a change in internal components or environmental interference. The description form is as follows: y is _out ＝αcos(f·t)t ₁ ≤t≤t ₂ Where α is the amplitude of the time-varying fault, f is the varying frequency parameter, t is the time variable, t ₁ ,t ₂ The time of occurrence and end of the fault. As shown in fig. 10 (b).

3) The stuck anomaly is represented by a fixed value of signal output, possibly no longer changing the element in response to the input signal, and possibly a loss of signal, and is described in the form of: y is _out ＝f _s Wherein f is _s The anomaly is a constant value, and as shown in fig. 10(c), the anomaly is constructed without human intervention, and is a true anomaly.

And 7: performing real-time prediction on the data after the standardization processing based on the trained model;

and step 8: and (4) calculating a prediction error, and if the prediction error exceeds the threshold value of the parameter, performing abnormal warning.

To evaluate the effectiveness of the proposed PA-LSTM (plus correlation parameters and attention mechanism), three models were added for comparison, respectively: P-LSTM (with correlation parameters), A-LSTM (with correlation mechanism but without correlation parameters), and LSTM (with no correlation parameters but without correlation mechanism), the accuracy of the four models for anomaly identification is shown in Table 4, and the error and threshold are shown in FIG. 11(a): PA-LSTM; FIG. 11(b): P-LSTM; FIG. 11(c): A-LSTM; FIG. 11(d): LSTM). The time-varying anomaly and the constant anomaly are artificially constructed anomalies, the four models of the time-varying anomaly can be identified, the constant anomaly PA-LSTM can be effectively identified as shown in fig. 11(a), and the A-LSTM can be effectively identified as shown in fig. 11(b), while the LSTM without an attention layer has higher prediction error, and in some cases, the P-LSTM is shown in fig. 11(b), and the LSTM is not capable of identifying the constant anomaly as shown in fig. 11 (d). And the false alarm rate of the LSTM is higher than that of the LSTM added with an attention layer (false alarm is the number of the originally normal LSTM but the abnormal LSTM is recognized).

The identification of true seizure anomalies by the four models is shown in FIG. 12(a): PA-LSTM; FIG. 12(b): P-LSTM; FIG. 12(c): A-LSTM; FIG. 12(d): LSTM), where seizure anomalies are most often severe and signal loss occurs. When the signal loss card is dead, the predicted values and the actual values of the four models have larger deviation, and the abnormity identification is realized. However, the system is a whole body, the seizure abnormality does not necessarily occur suddenly, and the abnormality occurs earlier based on the attitude system dynamics and kinematics, the angular velocity and other parameters. Data observation shows that three indexes of the magnetometer are normal for a period of time before signal loss, but other parameters have abnormal indications, such as abnormal fluctuation of parameters from 25 days in 3 months to 27 days in 3 months, rotational speed control of momentum wheel 1(ZK 052067) and rotational speed control of momentum wheel 3 (ZK052069), and large deviation of fine-sensitive 1 sun angle 1(ZK052084) and fine-sensitive 1 sun angle 2(ZK052085) occurs in the morning of 25 days in 3 months and lasts for a period of time. At this time, the A-LSTM and the LSTM without the relevant parameters are predicted according to a pure time sequence, no abnormity is found, however, the PA-LSTM and the P-LSTM with the relevant parameters predict a large error early warning in the morning of 25 morning days and continue a large error all the time. And the two times of seizure abnormity are ended up to 4 months and 3 days, and the prediction error is recovered to a normal and small level.

TABLE 4 comparison of anomaly identification accuracy

In the spacecraft attitude system anomaly detection method utilizing the multi-element multi-step prediction technology, the complex multi-element time sequence data can be monitored in real time, the self rule of multi-dimensional (self-settable) attitude parameters and the correlation among the parameters are comprehensively considered, and various anomalies such as constant anomaly, disturbance anomaly and mode mismatching can be accurately identified. In the process of constructing the abnormal detection model, the normal working historical data is used for training, the trained model and parameters are used for real-time data monitoring, the length of a detection window can be set automatically, the required calculation amount is small during monitoring, the efficiency is high, and the effect is good. The working principle of the spacecraft in each attitude mode is different, so that the internal structural features and the mutual relations of high-dimensional data are also different.

It should be understood that, although the steps in the flowchart of fig. 1 are shown in order as indicated by the arrows, the steps are not necessarily performed in order as indicated by the arrows. The steps are not performed in the exact order shown and described, and may be performed in other orders, unless explicitly stated otherwise. Moreover, at least a portion of the steps in fig. 1 may include multiple sub-steps or multiple stages that are not necessarily performed at the same time, but may be performed at different times, and the order of performance of the sub-steps or stages is not necessarily sequential, but may be performed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

In one embodiment, as shown in fig. 13, there is provided a spacecraft attitude system anomaly detection apparatus using a multivariate multistage prediction technique, including: a sample data obtaining module 200, a training data set constructing module 210, a standardization processing module 220, a neural network training module 230, an error threshold value calculating module 240, an actual data obtaining module 250, an actual prediction data obtaining module 260 and a system anomaly detecting module 270, wherein:

the system comprises a sample data acquisition module 200, a data processing module and a data processing module, wherein the sample data acquisition module 200 is used for acquiring sample data of a plurality of parameters of the spacecraft attitude system, and the sample data is a plurality of data which are continuously arranged in a time period by taking time as a sequence;

a training data set constructing module 210, configured to perform stationarity analysis and correlation analysis on each sample data, select a plurality of stationary state parameters and auxiliary parameters related to each stationary state parameter, and construct a training data set according to the sample data corresponding to the stationary state parameters and the auxiliary parameters;

a standardization processing module 220, configured to standardize the training data set to obtain a standard data set;

the neural network training module 230 is configured to construct a multi-element multi-step prediction neural network based on an attention mechanism and a long-short term memory neural network, and input the standard data set into the multi-element multi-step prediction neural network for training to obtain a trained multi-element multi-step prediction neural network, where the trained multi-element multi-step prediction neural network has a capability of predicting data of a time period after a parameter according to data of a time period before the parameter;

an error threshold calculation module 240, configured to calculate an error threshold of each steady-state parameter according to the predicted data and the real data of each steady-state parameter by using a 3 sigma rule;

the actual data acquisition module 250 is used for acquiring actual data of historical time of a plurality of parameters to be detected in the spacecraft attitude system and standardizing the actual data to obtain standardized actual data;

an actual prediction data obtaining module 260, configured to input the standardized actual data into a trained multivariate multi-step prediction neural network, so as to obtain real-time prediction data of each parameter to be detected in a next time period;

and the system anomaly detection module 270 is configured to obtain real data of each parameter to be detected, calculate an error between the real-time real data of each parameter to be detected and real-time predicted data at a corresponding time, and detect whether the current spacecraft attitude system is abnormal by comparing the error with an error threshold of the corresponding parameter.

For specific limitations of the spacecraft attitude system anomaly detection device using the multi-step prediction technology, reference may be made to the above limitations of the spacecraft attitude system anomaly detection method using the multi-step prediction technology, and details are not repeated here. All modules in the anomaly detection device for the spacecraft attitude system by utilizing the multivariate multistep prediction technology can be wholly or partially realized by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.

In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as shown in fig. 14. The computer device comprises a processor, a memory, a network interface, a display screen and an input device which are connected through a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to realize the anomaly detection method of the spacecraft attitude system by utilizing the multivariate multistep prediction technology. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, a key, a track ball or a touch pad arranged on the shell of the computer equipment, an external keyboard, a touch pad or a mouse and the like.

Those skilled in the art will appreciate that the architecture shown in fig. 14 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by hardware instructions of a computer program, which can be stored in a non-volatile computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. Any reference to memory, storage, database, or other medium used in the embodiments provided herein may include non-volatile and/or volatile memory, among others. Non-volatile memory can include read-only memory (ROM), Programmable ROM (PROM), Electrically Programmable ROM (EPROM), Electrically Erasable Programmable ROM (EEPROM), or flash memory. Volatile memory can include Random Access Memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus Direct RAM (RDRAM), direct bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

All possible combinations of the technical features in the above embodiments may not be described for the sake of brevity, but should be considered as being within the scope of the present disclosure as long as there is no contradiction between the combinations of the technical features.

The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent application shall be subject to the appended claims.

Claims

1. The method for detecting the abnormality of the spacecraft attitude system by utilizing the multivariate multistep prediction technology is characterized by comprising the following steps of:

constructing a multi-element multi-step prediction neural network based on an attention mechanism and a long-short term memory neural network, inputting the standard data set into the multi-element multi-step prediction neural network for training to obtain a trained multi-element multi-step prediction neural network, wherein the trained multi-element multi-step prediction neural network has the capability of predicting data of a time period after a parameter according to data of a time period before the parameter; inputting the standard data set into a multi-element multi-step prediction neural network for training, wherein the obtaining of the trained multi-element multi-step prediction neural network comprises the following steps: the long-term and short-term memory neural network comprises a plurality of processing layers which are connected in sequence, and each processing layer receives sample data at different moments as input data and outputs output data corresponding to the processing layer after corresponding processing; obtaining vector weights output by each time step of the processing layer based on attention mechanism matching according to the output data of each time step of the processing layer and the corresponding real data; after the vector weight output at each time step of the processing layer and the output data at each time step of the processing layer are subjected to weighted calculation, the vector weight and the output data at each time step of the processing layer are integrated into the final prediction calculation of the multi-element multi-step prediction neural network;

2. The method of claim 1, wherein the stationarity analyzing and correlation analyzing each of the sample data to select a plurality of stationary state parameters and auxiliary parameters related to each stationary state parameter comprises:

3. The method for detecting the anomaly in the spacecraft attitude system according to claim 1, wherein before the training of the multivariate multi-step predictive neural network, the method further comprises:

4. A spacecraft attitude system anomaly detection method according to claim 1, comprising, before inputting said normalized actual data into a trained multivariate multi-step predictive neural network:

5. A method as claimed in claim 1, wherein calculating the error threshold for each steady state parameter using 3 sigma law based on the predicted and actual data for each steady state parameter comprises: