Satellite fault detection method and device based on asynchronous autoregressive prediction model

Abstract

The application discloses a satellite fault detection method and device based on an asynchronous autoregressive prediction model, and the method comprises the following steps: acquiring satellite data to be detected; inputting first preset time characteristic data of the satellite data to be detected into a pre-trained autoregressive prediction model to obtain a predicted value of second preset time characteristic data, and calculating a reconstruction error of the predicted value and an actual value of the second preset time characteristic data; comparing the reconstruction error with a preset threshold value, and detecting fault data in the satellite data to be detected according to a comparison result. The application realizes rapid, accurate and robust automatic satellite anomaly detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, and particularly relates to a satellite fault detection method and device based on an asynchronous autoregressive prediction model. BACKGROUND

[0002] With the development of modern aerospace technology, satellites of various purposes have become an important part of national security, national life and scientific research. With the increasing complexity and precision of satellite structures, various possible faults can affect the normal operation of satellites and have a significant impact on public safety. Therefore, quickly and accurately detecting faults during satellite operation is a key fundamental problem.

[0003] In the field of satellite anomaly detection, the artificial threshold method is one of the earliest methods. If any parameter involved exceeds the safety range predefined by human experts, an anomaly will be reported. Although the artificial threshold method has good interpretability and computational efficiency, it cannot detect faults within the safety range, and this method can only detect some simple specific faults and cannot detect more complex fault types in more detail. In addition, data-driven methods, especially machine learning-based algorithms, can utilize large amounts of data to automatically identify fault types. K-nearest neighbor and decision tree algorithms use unsupervised training strategies to fit telemetry data and use training samples to determine fault thresholds. Support vector machine-based detection algorithms can mine potential information from normal and abnormal samples to improve detection accuracy. For example, Eskin et al. designed a clustering scheme and identified outliers as faults. Pajouh et al. combined dimensionality reduction and classification models for anomaly detection. George et al. used multi-class support vector machines. Compared with manual thresholds, anomaly detection based on machine learning technology combines the evaluation of observed parameters and identifies potential patterns. However, these methods may ignore the trend information of time input and cannot be well generalized to unseen anomaly types. When the distribution difference between abnormal data and normal data is not obvious, these methods cannot achieve fine-grained detection and obtain relatively poor accuracy, and lack the use of time series of telemetry data.

[0004] With the growth of data and computing facilities, deep learning algorithms have received extensive attention and achieved remarkable success in a large amount of data processing, and have become the most popular method in anomaly detection field. In recent years, therefore, some researchers have also tried to use neural networks for anomaly detection. Classification algorithms based on fully connected networks can achieve excellent detection results by using label information, but cannot handle unlabeled data situations. Recurrent neural networks (RNN) can be used for sequence modeling with hidden states. Long short-term memory networks (LSTM) introduce forget gates, input gates and output gates on the basis of conventional RNNs, improving the ability to model long-range relationships. They can handle both labeled and unlabeled data simultaneously and have been widely used in anomaly detection tasks with time input. However, existing methods still lack fine design for real satellite telemetry data.

[0005] In the case of satellite anomaly detection, the input is a time series, and the distribution of normal and abnormal data is extremely unbalanced, which makes it impossible to obtain an effective training model using conventional training methods. In addition, in practical applications, there may be types of anomalies in the training set that do not occur, making them unable to be detected in practical applications. SUMMARY

[0006] The present application aims to at least partially solve one of the technical problems in the related art.

[0007] To this end, the present application aims to solve the automatic and rapid anomaly detection of satellite data. Existing satellite operations usually require real-time manual monitoring to timely detect faults, and automatic anomaly detection methods based on classification models cannot be applied to asynchronous, non-memory time series satellite data. The present application realizes fast, accurate and robust automatic satellite anomaly detection by a satellite fault detection method based on an asynchronous autoregressive prediction model, which comprises a median filter preprocessing module, a non-uniform time encoding module, an autoregressive prediction model and an adaptive threshold selection module.

[0008] To achieve the above purpose, the present application proposes, in one aspect, a satellite fault detection method based on an asynchronous autoregressive prediction model, comprising:

[0009] acquiring satellite data to be detected;

[0010] inputting first predetermined time feature data of the satellite data to be detected into a pre-trained autoregressive prediction model to obtain a predicted value of second predetermined time feature data, and calculating a reconstruction error of the predicted value and the true value of the second predetermined time feature data;

[0011] comparing the reconstruction error with a preset threshold, and detecting fault data in the satellite data to be detected according to the comparison result.

[0012] The satellite fault detection method based on the asynchronous autoregressive prediction model according to the embodiment of the present application can further have the following additional technical features.

[0013] Further, in an embodiment of the present application, before the first preset time characteristic data of the satellite data to be detected is input into the pre-trained autoregressive prediction model, the method further comprises: obtaining a training data set, wherein the training data set comprises normal satellite data samples and abnormal satellite data samples; performing data conversion on the training data set by using non-uniform time encoding, training the autoregressive prediction model by using the normal satellite data samples after data conversion, obtaining the pre-trained autoregressive prediction model; and calculating the reconstruction error of the training data set according to the labels of the normal satellite data samples and the labels of the abnormal satellite data samples, and obtaining the threshold value by calculating the optimal Bayesian classification surface of the reconstruction error of the training data set.

[0014] Further, in an embodiment of the present application, the obtaining of the training data set comprises: obtaining multiple frames of data of initial satellite data samples at a third preset time; performing feature extraction on the multiple frames of data to obtain characteristic data at the third preset time; and performing preprocessing on the characteristic data at the third preset time by using a median filtering method to obtain the training data set.

[0015] Further, in an embodiment of the present application, the data conversion on the training data set by using non-uniform time encoding comprises: obtaining a timestamp sequence of the characteristic data of the training data set at a fourth preset time; obtaining relative information of the timestamp sequence and the fourth preset time by using a difference method; obtaining a problem function of the relative information by using a preset activation function, and querying the problem function by using query data for data conversion.

[0016] Further, in an embodiment of the present application, the training of the autoregressive prediction model by using the normal satellite data samples after data conversion to obtain the pre-trained autoregressive prediction model comprises: inputting characteristic data of the normal satellite data samples after data conversion at a fifth preset time into the autoregressive prediction model to obtain a sample characteristic data prediction value at the fifth preset time; training the autoregressive model based on the sample characteristic data prediction value at the fifth preset time and a true value to obtain the pre-trained autoregressive prediction model.

[0017] To achieve the above-mentioned purposes, another aspect of the present application provides a satellite fault detection device based on an asynchronous autoregressive prediction model, comprising:

[0018] a data acquisition module configured to acquire satellite data to be detected;

[0019] an error calculation module configured to input first preset time feature data of the satellite data to be detected into a pre-trained autoregressive prediction model to obtain a predicted value of second preset time feature data, and calculate a reconstruction error of the predicted value and a real value of the second preset time feature data;

[0020] a fault detection module configured to compare the reconstruction error with a preset threshold, and detect fault data in the satellite data to be detected according to a comparison result.

[0021] The satellite fault detection method and device based on the autoregressive prediction model of the application realize rapid, accurate and robust automatic satellite anomaly detection.

[0022] The technical effects of the application are as follows:

[0023] First, the satellite data is preprocessed to effectively eliminate outliers, then the data is asynchronously time-coded to maximize the preservation of original information, the autoregressive prediction model is used to accurately fit the distribution of normal satellite data, the accurate fault determination threshold is adaptively obtained based on prior knowledge, and finally the rapid, accurate, robust and automatic detection of satellite faults is realized, and the fault detection accuracy is more than 90%.

[0024] Additional aspects and advantages of the application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0025] The above and / or additional aspects and advantages of the application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which:

[0026] Figure 1 a flowchart of the satellite fault detection method based on the autoregressive prediction model according to an embodiment of the application;

[0027] Figure 2 a schematic diagram of the satellite fault detection based on the autoregressive prediction model according to an embodiment of the application;

[0028] Figure 3 a median filtering data preprocessing schematic diagram according to an embodiment of the application;

[0029] Figure 4 a structure schematic diagram of the satellite fault detection device based on the autoregressive prediction model according to an embodiment of the application. DETAILED DESCRIPTION

[0030] It should be noted that the embodiments and features in the application can be combined with each other without conflict. The application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0031] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0032] The satellite fault detection method and device based on an asynchronous autoregressive prediction model according to the embodiments of the present application are described below with reference to the drawings.

[0033] Figure 1 is a flowchart of the satellite fault detection method based on an asynchronous autoregressive prediction model according to an embodiment of the present application.

[0034] As shown in Figure 1 , the method includes but is not limited to the following steps:

[0035] S1, acquiring satellite data to be detected.

[0036] Specifically, the satellite data acquired in the embodiments of the present application is initial satellite data, which is detected by the method of the embodiments of the present application.

[0037] S2, inputting the first preset time feature data of the satellite data to be detected into a pre-trained autoregressive prediction model to obtain a predicted value of the second preset time feature data, and calculating the reconstruction error of the predicted value and the true value of the second preset time feature data.

[0038] It can be understood that before detecting new satellite data, the trained model and the threshold value need to be obtained by the method of the embodiments of the present application.

[0039] As shown in Figure 2 , first, the data processing method of median filtering is used to eliminate abnormal outliers in the original data.

[0040] Specifically, for the feature of any time t i , K frames of data before and after it are selected, and the median of the K frames of data is taken as the feature of t i . As shown in Figure 3 , the median filtering method can effectively filter out outliers while perfectly preserving the step change mode in the signal.

[0041] It can be understood that there are a large number of non-anomalous outliers in the satellite feature data, and due to the high-frequency data in the original features, the application uses median filtering for data preprocessing to prevent the introduction of excessive low-frequency noise information in the preprocessing process.

[0042] Further, as shown in Figure 2 , satellite data transformation is performed using non-uniform time encoding. A general autoregressive model needs to use an interpolation algorithm to uniformize the time interval, but due to the possibility of high-frequency information in satellite data changes, the interpolation algorithm is easy to introduce additional low-frequency information to destroy the original data pattern, and the embodiment of the application maximizes the preservation of original information through non-uniform time encoding.

[0043] Specifically, let the satellite feature at time t be x_t∈R^d, where d is the data dimension. For the feature block X corresponding to the timestamp sequence δ=(t1,t2,…,t T )∈R T , we first use the difference method to only retain the relative information of the time points Then use the tanh activation function to prevent numerical instability δ=tanhδ, and to mark the time point to be predicted, we additionally increase the query item t p , so that the final input data is:

[0044]

[0045] The prediction value of the model for t p is D pre =f θ (D in )∈R d .

[0046] Further, as shown in Figure 2 , a self-regressive prediction model is used to model the satellite data distribution. Since the satellite anomaly detection task often cannot obtain real abnormal samples, and the features between adjacent time points conform to the Markov assumption, the application proposes to use a self-regressive prediction model to solve the prediction problem of unknown abnormal situations.

[0047] Specifically, the application uses a self-regressive model f θ , intercepts the feature block X=[x1,x2,…,x T ]∈R T×d of T consecutive time points of the satellite, and tries to make the model fit the feature of the T+1 time point. A training set with only normal samples is constructed, and the network is trained by minimizing the mean square error. The specific formula is as follows:

[0048]

[0049] Make it possible for the model to learn the dynamic mode of the satellite in the normal state.

[0050] Further, as Figure 2 Indicated, considering prior information, an adaptive threshold selection method is used to more accurately determine the abnormal threshold.

[0051] Specifically, if the abnormal sample has appeared in the data, we can directly use the abnormal label; if there is no abnormal sample in the data, we can generate abnormal samples (for example, gradual drift anomaly or sudden step anomaly) according to the characteristics of the satellite. According to the existing sample label (normal sample y i =1 and abnormal sample y i =-1), the reconstruction error e i of all samples is calculated, and the threshold is calculated by calculating the optimal Bayesian classification surface. The specific calculation formula is:

[0052]

[0053] Among them,

[0054]

[0055] Understandably, the feature data of a certain moment of the satellite data to be detected is input into the pre-trained autoregressive prediction model, and the predicted value of the next moment feature data is obtained, so as to calculate the reconstruction error of the predicted value and the true value.

[0056] S3, compare the reconstruction error with the preset threshold, and detect the fault data in the satellite data to be detected according to the comparison result.

[0057] Understandably, the model is used to determine the fault of the satellite data.

[0058] Specifically, in the test phase, the prediction model relies on the reconstruction error to judge whether the new sample is abnormal. The specific method is that, given the features D in of the previous T moments of the time stamp to be judged, the model is used to predict the t p th moment feature The reconstruction error of the predicted value and the true value is calculated For a given threshold value ∈, if is greater than the threshold value, it is judged that the t p th moment data is abnormal, otherwise it is judged to be normal. Therefore, the selection of the threshold value is very important.

[0059] In the embodiment of the application, adaptive threshold selection is used for fault determination, and the use of normal sample statistics as a threshold does not introduce priori. The application proposes an adaptive threshold construction method based on prior knowledge, which can obtain a more effective classifier.

[0060] The satellite fault detection method based on the asynchronous autoregressive prediction model according to the embodiment of the present application realizes rapid, accurate and robust automatic satellite anomaly detection.

[0061] To achieve the above-mentioned embodiments, as Figure 4 shown, the embodiment also provides a satellite fault detection device 10 based on an asynchronous autoregressive prediction model, which comprises a data acquisition module 100, an error calculation module 200 and a fault detection module 300.

[0062] The data acquisition module 100 is configured to acquire satellite data to be detected.

[0063] The error calculation module 200 is configured to input the first preset time feature data of the satellite data to be detected into a pre-trained autoregressive prediction model to obtain a predicted value of the second preset time feature data, and calculate a reconstruction error of the predicted value and the real value of the second preset time feature data.

[0064] The fault detection module 300 is configured to compare the reconstruction error with a preset threshold, and detect fault data in the satellite data to be detected according to the comparison result.

[0065] Further, before the error calculation module 200, a model training module is further included, and the model training module comprises:

[0066] The training data acquisition module is configured to acquire a training data set, wherein the training data set comprises normal satellite data samples and abnormal satellite data samples.

[0067] The data conversion training module is configured to convert the training data set by using non-uniform time encoding, train the autoregressive prediction model by using the normal satellite data samples after data conversion, and obtain a pre-trained autoregressive prediction model.

[0068] The threshold calculation and determination module is configured to calculate the reconstruction error of the training data set according to the labels of the normal satellite data samples and the labels of the abnormal satellite data samples, and obtain the threshold by calculating the optimal Bayesian classification surface of the reconstruction error of the training data set.

[0069] Further, the training data acquisition module is further configured to:

[0070] Acquire multiple frames of data of the initial satellite data samples at a third preset time.

[0071] Feature extraction is performed on the multiple frames of data to obtain feature data at the third preset time.

[0072] The feature data at the third preset time is preprocessed by using a median filtering method to obtain the training data set.

[0073] Further, the data conversion training module is further used for:

[0074] obtaining a timestamp sequence of the feature data of the training data set at the fourth preset time point;

[0075] obtaining relative information of the timestamp sequence and the fourth preset time point by using a difference method;

[0076] obtaining a problem function of the relative information by using a preset activation function, and querying the problem function by using the query data to perform data conversion.

[0077] Further, the data conversion training module is further used for:

[0078] inputting the feature data of the normal satellite data sample at the fifth preset time point into the autoregressive prediction model to obtain a sample feature data prediction value at the fifth preset time point;

[0079] training the autoregressive model based on the sample feature data prediction value at the fifth preset time point and the real value to obtain a pre-trained autoregressive prediction model.

[0080] According to the satellite fault detection device based on the asynchronous autoregressive prediction model, fast, accurate and robust automatic satellite anomaly detection is realized.

[0081] In addition, the terms "first", "second", "third", "fourth", "fifth", "sixth", "seventh" and "eighth" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second", "third", "fourth", "fifth", "sixth", "seventh" and "eighth" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise explicitly specified.

[0082] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the skilled in the art can combine and combine the different embodiments or examples described in the present application and the features of the different embodiments or examples without contradiction.

[0083] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and that changes, modifications, substitutions and variations can be made by those skilled in the art without departing from the scope of the present application.

Claims

1. A satellite fault detection method based on an asynchronous autoregressive prediction model, characterized in that, The method comprises the following steps: acquiring satellite data to be detected; inputting first preset time feature data of the satellite data to be detected into a pre-trained autoregressive prediction model to obtain a predicted value of second preset time feature data, and calculating a reconstruction error of the predicted value and a real value of the second preset time feature data; comparing the reconstruction error with a preset threshold value, and detecting fault data in the satellite data to be detected according to a comparison result; before the step of inputting the first preset time feature data of the satellite data to be detected into the pre-trained autoregressive prediction model, the method further comprises the following steps: acquiring a training data set, wherein the training data set comprises normal satellite data samples and abnormal satellite data samples; performing data conversion on the training data set by using a non-uniform time encoding, training an autoregressive prediction model by using the normal satellite data samples after data conversion to obtain the pre-trained autoregressive prediction model; and calculating a reconstruction error of the training data set according to labels of the normal satellite data samples and labels of the abnormal satellite data samples, and obtaining the threshold value by using an optimal Bayesian classification surface of the reconstruction error of the training data set; the step of acquiring the training data set comprises the following steps: acquiring multiple frames of data of initial satellite data samples at a third preset time; performing feature extraction on the multiple frames of data to obtain feature data at the third preset time; performing preprocessing on the feature data at the third preset time by using a median filtering method to obtain the training data set.

2. The method of claim 1, wherein, the step of performing data conversion on the training data set by using a non-uniform time encoding comprises the following steps: acquiring a timestamp sequence of feature data of the training data set at a fourth preset time; obtaining relative information of the timestamp sequence and the fourth preset time by using a difference method; obtaining a problem function of the relative information by using a preset activation function, and querying the problem function by using query data to perform data conversion.

3. The method of claim 1, wherein, the step of training the autoregressive prediction model by using the normal satellite data samples after data conversion to obtain the pre-trained autoregressive prediction model comprises the following steps: inputting feature data of the normal satellite data samples after data conversion at a fifth preset time into the autoregressive prediction model to obtain a predicted value of sample feature data at the fifth preset time; training the autoregressive prediction model based on the predicted value of sample feature data at the fifth preset time and a real value to obtain the pre-trained autoregressive prediction model.

4. A satellite fault detection apparatus based on an asynchronous autoregressive prediction model, characterized by, The method comprises the following steps: a data acquisition module is configured to acquire satellite data to be detected; an error calculation module is configured to input first preset time feature data of the satellite data to be detected into a pre-trained autoregressive prediction model to obtain a predicted value of second preset time feature data, and calculate a reconstruction error of the predicted value and a real value of the second preset time feature data; a fault detection module is configured to compare the reconstruction error with a preset threshold value, and detect fault data in the satellite data to be detected according to a comparison result; before the error calculation module, a model training module is further included, and the model training module comprises the following steps: The training data acquisition module is configured to acquire a training data set, wherein the training data set includes normal satellite data samples and abnormal satellite data samples. The data conversion training module is configured to perform data conversion on the training data set by using non-uniform time encoding, train an autoregressive prediction model by using the normal satellite data samples after data conversion, and obtain the pre-trained autoregressive prediction model. The threshold calculation and determination module is configured to calculate reconstruction errors of the training data set according to labels of the normal satellite data samples and labels of the abnormal satellite data samples, and obtain the threshold by calculating an optimal Bayesian classification surface of the reconstruction errors of the training data set. The training data acquisition module is further configured to: acquire multiple frames of data of an initial satellite data sample at a third preset time; perform feature extraction on the multiple frames of data to obtain feature data at the third preset time; perform preprocessing on the feature data at the third preset time by using a median filtering method to obtain the training data set.

5. The apparatus of claim 4, wherein, The data conversion training module is configured to: acquire a timestamp sequence of the feature data of the training data set at a fourth preset time; obtain relative information of the timestamp sequence and the fourth preset time by using a difference method; obtain a problem function of the relative information by using a preset activation function, and query the problem function by using query data to perform data conversion.

6. The apparatus of claim 4, wherein, The data conversion training module is further configured to: input feature data of the normal satellite data samples after data conversion at a fifth preset time into the autoregressive prediction model to obtain a sample feature data prediction value at the fifth preset time; train the autoregressive prediction model based on the sample feature data prediction value at the fifth preset time and an actual value to obtain the pre-trained autoregressive prediction model.

Images