Satellite multi-dimensional telemetry sequence anomaly detection model construction method and device, anomaly detection method and device

Abstract

The invention relates to a method and device for constructing a satellite multi-dimensional telemetry sequence anomaly detection model, and an anomaly detection method and device, and relates to the field of data processing and detection. In view of the technical problem in the prior art that the existing multi-dimensional time series anomaly detection algorithm is not good at modeling the complex interactions and dependencies between telemetry data and control commands, the technical solution provided by the present invention is: a method for constructing a satellite multi-dimensional telemetry sequence anomaly detection model, the method comprising: a step of collecting telemetry data; a step of mapping the data into time-correlated latent variables; a step of obtaining reconstructed input data based on the time-correlated latent variables; a step of obtaining spatially correlated latent variables based on the reconstructed input data; a step of obtaining the difference between telemetry data and reconstructed data based on the spatially correlated latent variables. The satellite multi-dimensional telemetry sequence anomaly detection method provided by the present invention is suitable for application in the work of spacecraft telemetry sequence anomaly detection.

Description

Technical Field

[0001] It relates to the field of data processing and detection, and specifically to anomaly detection in spacecraft telemetry sequences. Background Art

[0002] In recent years, deep learning models have become a popular method for anomaly detection in multidimensional telemetry sequences. In this regard, Kyle et al. proposed a non-parametric dynamic threshold method based on long short-term memory networks (LSTMs) for detecting spacecraft anomalies. This method uses the memory and sequence modeling capabilities of LSTM networks to capture normal patterns in telemetry data and identifies abnormal behaviors by comparing the observed telemetry data with the predicted output of the LSTM model. However, this method is mainly designed for modeling temporal dependencies and ignores the correlation between spatial features.

[0003] To address this problem, a graph learning with Transformer (GTA) anomaly detection framework is proposed, which includes automatic learning of graph structures, graph convolutions, and using a Transformer-based architecture to model the temporal dependencies of multi-dimensional telemetry sequences. However, when the spatial correlation is complex and nonlinear, the learned graph structure cannot detect subtle anomalies well and cannot effectively model the dependencies between spatial features.

[0004] Another approach called Attention Temporal Convolutional Network (ATCN) focuses on detecting anomalies at the entity level. It uses temporal convolutional neural networks and dynamic graph attention techniques to model dependencies between spatial features. However, this approach lacks explicit low-dimensional temporal embeddings, and building a model for each telemetry entity also leads to weak detection capabilities for small anomalies involving each channel.

[0005] Another unsupervised method is called InterFusion, which is based on variational autoencoders (VAE) and Markov chain Monte Carlo (MCMC). This method models the normal patterns within time series data by learning a robust multi-dimensional time series representation. This method has high accuracy and robustness for continuous data types. However, when dealing with mixed data types such as telemetry data and telecommands, this method is not suitable for modeling complex, nonlinear correlations between spatial features.

[0006] For satellites in orbit, telemetry data is collected in time sequence. Complex system design affects system performance, making telemetry sequences have spatial correlation characteristics. Changes in spatial and temporal correlations may lead to anomalies in telemetry parameters. However, previous research works mainly focus on modeling the temporal or spatial relationships of telemetry data, which limits the model's ability to learn telemetry data patterns and detect anomalies.

[0007] As a complex system, telemetry instructions usually have complex long-term dependencies that affect telemetry data. Telemetry instructions are usually remote control instructions represented by a binary sequence of 0s and 1s, while telemetry data is a continuous reflection of system status and performance collected by sensors. Changes in telemetry data may be triggered or regulated by specific telemetry instructions, and there is usually a delay in the interaction between instructions and telemetry sequence data.

[0008] Therefore, when designing telemetry anomaly detection methods, it is necessary to comprehensively consider temporal dependency, spatial correlation, and the interaction between instructions and telemetry sequences. A potential approach is to combine long short-term memory networks (LSTM) and graph convolutional neural networks (GCN), using LSTM networks to model temporal dependencies and GCN networks to model the relationship between spatial features. This approach can capture both temporal dependencies and spatial correlations, and can flexibly adapt to different types of data in complex systems.

[0009] Another direction is to combine variational autoencoders (VAE) and graph attention mechanisms to build a model that can handle mixed data types and detect complex spatial correlations. VAE can learn low-dimensional representations of time series data, while graph attention mechanisms can dynamically adjust the attention between different spatial features.

[0010] In summary, for the problem of telemetry anomaly detection, it is necessary to comprehensively consider time dependency, spatial correlation, and the interaction between instructions and telemetry sequences when designing the model. Combining different deep learning techniques, such as LSTM, GCN, VAE, and graph attention mechanism, more powerful, accurate, and robust telemetry anomaly detection models can be built. The development of these models will help improve the reliability and safety of satellite operations and promote the performance optimization of spacecraft and their related systems.

[0011] However, existing multi-dimensional time series anomaly detection algorithms are not good at modeling the complex interactions and dependencies between telemetry data and control commands. Summary of the invention

[0012] In view of the technical problem in the prior art that the existing multidimensional time series anomaly detection algorithm is not good at modeling the complex interactions and dependencies between telemetry data and control commands, the technical solution provided by the present invention is as follows:

[0013] A method for constructing a satellite multi-dimensional telemetry sequence anomaly detection model, the method comprising:

[0014] Steps to collect telemetry data;

[0015] The step of mapping the data into time-dependent latent variables;

[0016] According to the latent variables of the time correlation, a step of reconstructing input data is obtained;

[0017] A step of obtaining latent variables of spatial correlation according to the reconstructed input data;

[0018] The step of obtaining the difference between the telemetry data and the reconstructed data according to the latent variable of the spatial correlation.

[0019] Furthermore, a preferred embodiment is provided, in which the latent variables of the time correlation are mapped through a convolutional network.

[0020] Furthermore, a preferred implementation is provided, in which a step of reconstructing input information is obtained through a deconvolution network.

[0021] Furthermore, a preferred implementation is provided, in which the latent variables of the time correlation are mapped back to the original data space through a deconvolution network to obtain reconstructed input information.

[0022] Furthermore, a preferred implementation is provided to obtain latent variables of spatial correlation through GRU and attention mechanism.

[0023] Based on the same inventive concept, the present invention also provides a device for constructing a satellite multi-dimensional telemetry sequence anomaly detection model, the device comprising:

[0024] Module for collecting telemetry data;

[0025] A module for mapping the data into time-dependent latent variables;

[0026] According to the latent variables of the time correlation, a module for reconstructing input data is obtained;

[0027] A module for obtaining latent variables of spatial correlation according to the reconstructed input data;

[0028] A module for obtaining the difference between the telemetry data and the reconstructed data based on the latent variables of the spatial correlation.

[0029] Based on the same inventive concept, the present invention also provides a method for detecting anomalies in a satellite multi-dimensional telemetry sequence, the method comprising:

[0030] Steps for collecting the measured telemetry data sequence;

[0031] According to the satellite multi-dimensional telemetry sequence anomaly detection model construction method, a step of obtaining the difference between the telemetry data and the reconstructed data of the sequence.

[0032] Based on the same inventive concept, the present invention also provides a satellite multi-dimensional telemetry sequence anomaly detection device, the device comprising:

[0033] A module for collecting the measured telemetry data sequence;

[0034] According to the satellite multi-dimensional telemetry sequence anomaly detection model construction device, a module for obtaining the difference between the telemetry data and the reconstructed data of the sequence is obtained.

[0035] Based on the same inventive concept, the present invention also provides a computer storage medium for storing a computer program. When the program is read by a computer, the computer executes the method described.

[0036] Based on the same inventive concept, the present invention also provides a computer, including a processor and a storage medium. When the processor reads a computer program stored in the storage medium, the computer executes the method described.

[0037] Compared with the prior art, the technical solution provided by the present invention is beneficial in that:

[0038] The invention provides a satellite multi-dimensional telemetry sequence anomaly detection method, wherein a hierarchical variational autoencoder with two random latent variables is intended to be used for modeling the spatial and temporal correlation of the multi-dimensional telemetry sequence.

[0039] The satellite multi-dimensional telemetry sequence anomaly detection method provided by the present invention integrates a self-attention mechanism based on a GRU neural network to enhance the model's ability to process the long-term dependency of telemetry data on remote control commands.

[0040] The satellite multi-dimensional telemetry sequence anomaly detection method provided by the present invention is suitable for application in the work of spacecraft telemetry sequence anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic diagram of the structure of a satellite multi-dimensional telemetry sequence anomaly detection model provided in implementation mode 1;

[0042] Figure 2 A flowchart of a satellite multi-dimensional telemetry sequence anomaly detection method provided for implementation mode seven;

[0043] Figure 3 This is a schematic diagram of the comparison between the actual abnormal sequence and the abnormal detection result mentioned in the eleventh embodiment;

[0044] Among them, (a) is the actual anomaly sequence, and (b) is the anomaly detection result. DETAILED DESCRIPTION

[0045] In order to make the advantages and benefits of the technical solution provided by the present invention more clearly reflected, the technical solution provided by the present invention is now further described in detail in conjunction with the accompanying drawings:

[0046] Implementation Method 1: Combination Figure 1This embodiment describes a method for constructing a satellite multi-dimensional telemetry sequence anomaly detection model. The method includes:

[0047] Steps to collect telemetry data;

[0048] The step of mapping the data into time-dependent latent variables;

[0049] According to the latent variables of the time correlation, a step of reconstructing input data is obtained;

[0050] A step of obtaining latent variables of spatial correlation according to the reconstructed input data;

[0051] The step of obtaining the difference between the telemetry data and the reconstructed data according to the latent variable of the spatial correlation.

[0052] Specific:

[0053] The telemetry data x is first processed by a convolutional network. The convolutional layer can extract the local features of the input data and map x to a hidden representation z1 through a combination of convolution and activation functions.

[0054] Input z1 into the deconvolution network, which maps the hidden representation z1 back to the original data space through the reverse convolution operation to obtain the reconstructed input d. The reconstructed input d is an approximate reconstruction of the input data x.

[0055] The reconstructed input d is passed as input to the SRNN-like network. The SRNN network processes sequence data by using techniques such as Gated Recurrent Unit (GRU) and Attention Mechanism.

[0056] GRU (Gated Recurrent Unit): GRU is a gated recurrent unit that can capture long-term dependencies in sequence data. GRU controls the transmission and retention of information and determines the hidden state of the current time step by updating and resetting gates.

[0057] Attention Mechanism: The attention mechanism can adaptively select important information in sequence data. It combines the reconstructed input d and the previous hidden state by calculating the weights at different time steps in order to focus on information with higher importance in the SRNN network.

[0058] Processing to obtain z2: The SRNN network, through calculations at multiple time steps, combined with GRU and attention mechanisms, gradually processes and reconstructs the input d and outputs the hidden representation z2. In this way, z2 can be used for subsequent detection tasks.

[0059] in,

[0060] For the original data X∈R M×N , each row x i is called a feature, and each column x t is called an observation.

[0061] Hidden variable z1=g(x)∈R m×w′ , g is a number of one-dimensional convolutions, m is the number of features, and w' is the length of the window after convolution.

[0062] d=f(z1)∈R m×w , and f is the corresponding one-dimensional deconvolution layer.

[0063] Hidden variable z2∈R m′×w , is the embedded representation compressed along the feature dimension on the reconstructed d, and m' is the feature dimension after compression.

[0064] Latent variables represent the potential characteristics or representation of the input data.

[0065] Hidden variable z1 of time correlation;

[0066] Latent variable z2 of spatial correlation;

[0067] After obtaining the latent variable z1 of time correlation and the latent variable z2 of space correlation, the difference or loss between the original data and the reconstructed data is calculated through the reconstruction error function (L) to achieve the minimum or optimize it to the minimum.

[0068]

[0069] in, represents the expected logarithmic difference between the probability of data in the original distribution and the approximate distribution, p θ represents the probability distribution of the Kullback-Leibler divergence, e represents the reconstructed input required by the generative network (pnet), and q φ represents the upper approximation distribution of the Kullback-Leibler divergence.

[0070] The model is to achieve better results through training, and the difference or loss between the original data and the reconstructed data is calculated through the reconstruction error function (formula L) to achieve the minimum or optimize to the minimum. For example, there are many groups of weights, and I am not sure which group is suitable for my model. Through training, let the model find the right group by itself so that the loss of formula L is smaller, that is, the model effect is better.

[0071] In anomaly detection problems, we want to identify unusual data points that do not conform to normal patterns. A common approach is to use a reconstruction model that attempts to generate a close copy or reconstruction from the original data and measures the degree of anomaly by comparing the difference between the original and the reconstructed data.

[0072] The reconstruction probability refers to the probability distribution of the data points generated by the reconstruction model in the model. For normal data points, they usually have a higher reconstruction probability because the model can reconstruct these data points well. For abnormal data points, they may be very different from the normal pattern, making it difficult for the model to accurately reconstruct them, so they have a lower reconstruction probability.

[0073] The generator network (pnet in the figure) is the part of HVAE that is responsible for generating samples from the latent space. The generator network receives random vectors from the latent variable space (usually following a standard normal distribution) as input and transforms them into generated samples with similar characteristics to the original data through a series of mapping and transformation operations. The goal of the generator network is to learn the distribution of the data so that it can generate new samples with characteristics similar to the training data.

[0074] The variational posterior network (qnet in the figure) is also an important part of HVAE, which is responsible for mapping the input data to the distribution in the latent space. The variational posterior network receives samples from the input data as input and converts them into the distribution parameters of the latent variables through a series of mapping and transformation operations. The goal of the variational posterior network is to learn the distribution of data in the latent space so that the latent variables can be inferred and reconstructed for the given input data.

[0075] The generative network and the variational posterior network are interrelated. In HVAE, the generative network and the variational posterior network are trained together by minimizing the KL divergence in variational inference. The generative network learns the ability to generate samples by minimizing the reconstruction error (usually the error between the reconstructed sample and the original data), while the variational posterior network learns the latent variable distribution of the data by minimizing the KL divergence. The training of these two networks is carried out in coordination, allowing HVAE to learn the ability to generate samples and infer latent variables at the same time.

[0076] Model architecture part:

[0077] The telemetry sequences of spacecraft exhibit complex interdependencies, including temporal correlations (e.g., periodicity) and variable correlations (linear or nonlinear relationships within the variables of each time period entity). A hierarchical variational autoencoder with two random latent variables can model the normal patterns of telemetry sequences. On this basis, we integrate a self-attention mechanism based on a GRU neural network, which enables the model to adaptively learn important features in the telemetry data, thereby deriving latent variables that characterize spatial and temporal correlations. The model architecture is shown in Figure 2. Figure 1 Specifically, the generated model (pnet) can be expressed as:

[0078] p θ (x,z2,z1)=p θ (x|z2,z1)p θ (z2|z1)p θ (z1);

[0079] By jointly training random latent variables, z1 and z2 learn temporal embedding or spatial feature embedding. The hierarchical structure of the layered variational autoencoder is used to transform the original input x into a low-dimensional representation z (z2|z1 hierarchy).

[0080] The purpose of this hierarchy is to enable z2 to combine with z1 to capture the learned temporal information, rather than learning spatial feature embeddings independently. Specifically, for the temporal embedding layer, z1 = g(x)∈R m×w′ , g is a number of one-dimensional convolutions, m is the number of features, and w′ is the length of the window after convolution. Define d = f(z1)∈R m×w And f is the corresponding one-dimensional deconvolution layer. The spatial feature embedding z2∈R m′×w is the embedded representation compressed along the feature dimension on the reconstructed d, and m′ is the compressed feature

[0081] The variational autoencoder VAE is used to pre-train d to ensure its initial reconstructability in the early stages of model training. Figure 1 As shown, the variational posterior (qnet) can be expressed as:

[0082] q φ (z2,z1|x)=q φ (z2|z1)q φ (z1|x);

[0083] For each time instant t, the input value d t and c t+1 (The deterministic state derived by GRU at time t+1) is passed to a reverse recurrent GRU unit to generate the hidden variable c t :

[0084]

[0085] Finally, in order to well characterize the long-term dependencies in the input sequence, at each time t, c t is used as the input of the self-attention mechanism. Thus, the attention weights of different positions in the input sequence are obtained:

[0086] a t =softmax(w2 tanh(w1c t +b1);

[0087] Where w1, w2, and b1 represent the weights and biases of the corresponding fully connected layers. In particular, the parameters of the one-dimensional deconvolution layer (used to derive d, e) are shared between the generator network and the variational posterior network, so that the “reconstructed input” and the resulting temporal information can be shared, thereby improving the training effect.

[0088] Implementation method 2: This implementation method further limits the method for building a satellite multi-dimensional telemetry sequence anomaly detection model provided in implementation method 1, and maps the latent variables of the time correlation through a convolutional network.

[0089] Implementation method three: This implementation method further limits the method for building a satellite multi-dimensional telemetry sequence anomaly detection model provided in implementation method one, and obtains the step of reconstructing input information through a deconvolution network.

[0090] Implementation method 4: This implementation method further limits the satellite multi-dimensional telemetry sequence anomaly detection model construction method provided in implementation method 3. The time-correlated latent variables are mapped back to the original data space through a deconvolution network to obtain reconstructed input information.

[0091] Implementation method 5: This implementation method further limits the method for building a satellite multi-dimensional telemetry sequence anomaly detection model provided in implementation method 1. Through GRU and attention mechanism, latent variables of spatial correlation are obtained.

[0092] Embodiment 6: This embodiment provides a device for constructing a satellite multi-dimensional telemetry sequence anomaly detection model, the device comprising:

[0093] Module for collecting telemetry data;

[0094] A module for mapping the data into time-dependent latent variables;

[0095] According to the latent variables of the time correlation, a module for reconstructing input data is obtained;

[0096] A module for obtaining latent variables of spatial correlation according to the reconstructed input data;

[0097] A module for obtaining the difference between the telemetry data and the reconstructed data based on the latent variables of the spatial correlation.

[0098] Implementation Method 7: Combination Figure 2 This embodiment describes a method for detecting anomalies in a satellite multi-dimensional telemetry sequence. The method comprises:

[0099] Steps for collecting the measured telemetry data sequence;

[0100] According to the method provided in embodiment 1, a step is provided for obtaining the difference between the telemetry data and the reconstructed data of the sequence.

[0101] Specific:

[0102] Training and inference steps:

[0103] The SGVB algorithm is used to optimize the evidence lower bound (ELBO), where D KL The divergence is the Kullback-Leibler divergence. The mathematical formula is as follows:

[0104]

[0105] After introducing the reconstructed inputs d and e, the model optimization objective becomes the following:

[0106]

[0107] In online detection, our goal is to use a given sliding window {x t-w+1 , ..., x t} to detect timestamp x t Is it an exception and calculate the reconstruction probability This is widely used in anomaly detection. The last data x in the window t The reconstruction probability of is used as the anomaly score at time t.

[0108] Abnormal detection judgment process:

[0109] The trained model is used to determine the time step x t Is the observed value on abnormal to calculate the reconstruction probability As its anomaly score. Specifically, select {x t-w+1 , ..., x t} as a sliding window, that is, x t and its previous w-1 consecutive observations as input to detect anomalies at time t. The last data (x t ) is used as the anomaly score, which can speed up the response to anomaly detection. t The abnormal score is recorded as St , so S t =log(p θ (x t |z t-W+1：t )). In practice, anomalies are rarely independent, but rather continuous abnormal segments. Therefore, the point-adjust approach is mainly used as the evaluation strategy. Specifically, if the model can detect any subset of abnormal segments in the telemetry data, i.e., trigger an abnormal response, then all observations of the abnormal segment of the data are considered to be correctly detected. At the same time, points outside the abnormal segment are processed as usual. Formally, if S t If the observed value x is higher than the predefined threshold, t is marked as an exception, otherwise x t is normal. Figure 3 A judgment example of abnormality detection is described.

[0110] Embodiment 8: This embodiment provides a satellite multi-dimensional telemetry sequence anomaly detection device, the device comprising:

[0111] A module for collecting the measured telemetry data sequence;

[0112] According to the device provided in embodiment 6, a module is provided for obtaining the difference between the telemetry data and the reconstructed data of the sequence.

[0113] Embodiment 9: This embodiment provides a computer storage medium for storing a computer program. When the program is read by a computer, the computer executes the method provided in Embodiments 1 to 5 or 7.

[0114] Embodiment 10: This embodiment provides a computer, including a processor and a storage medium. When the processor reads a computer program stored in the storage medium, the computer executes the method provided in embodiments 1 to 5 or 7.

[0115] Implementation Method XI: Combination Figure 3 This embodiment is described. This embodiment is an example to verify the method provided in the seventh embodiment. Specifically:

[0116] Table 1 Parameter configuration

[0117]

[0118] As shown in Table 1, the sliding window size is set to 100 and the feature embedding M′ is set to 2. In addition, the dropout regularization strategy is used to prevent overfitting, and the dropout rate is 0.1. The model is optimized using the Adam algorithm with a learning rate of 1e -3, the batch size is set to 256, and the training iterations last for 20 epochs. RELU is used as the activation function for all layers except the linear layer, and the last 30% of the data in the training set is taken as the validation set. An early stopping strategy is used in training to prevent the model from overfitting.

[0119] Table 2 Performance of the compared methods

[0120]

[0121] As shown in Table 2, it can be seen that the designed model achieves the best F1 scores of 0.9388 and 0.9281 on the SMAP dataset and MSL dataset. Specifically, we can make the following analysis.

[0122] The proposed method shows superior performance compared to other baselines, with differences of 0.0132-0.0897 on the SMAP dataset and 0.0026-0.0371 on the MSL dataset. This makes it attractive enough to handle anomaly detection in multi-dimensional telemetry data. However, for some special datasets or anomalies, the proposed method may not achieve the best performance. Specifically, when the abnormal pattern is very similar or overlapping with the normal pattern, it may be difficult for the algorithm to accurately distinguish the abnormal behavior. When the abnormal pattern changes are small, the effectiveness of the algorithm may be limited. In addition, if there are anomalies in the training data, overfitting may occur, which may cause the learned feature embedding to deviate significantly from the normal embedding.

[0123] The technical solution provided by the present invention is further described in detail above through several specific implementation modes in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific implementation modes described above are not intended to be used as limitations on the present invention. Any reasonable modification and improvement of the present invention, reasonable combination of implementation modes and equivalent substitution within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for constructing anomaly detection model for satellite multi-dimensional telemetry sequences, characterized in that: The method comprises: Steps to collect telemetry data; The step of mapping the data into time-dependent latent variables; According to the latent variables of the time correlation, a step of reconstructing input data is obtained; A step of obtaining latent variables of spatial correlation according to the reconstructed input data; A step of obtaining a difference between the telemetry data and the reconstructed data according to the latent variables of the spatial correlation; For the data X∈R M×N , each row x i is called a feature, and each column x t is called an observation; Hidden variable z1=g(x)∈R m×w′ , g is a number of one-dimensional convolutions, m is the number of features, and w' is the length of the window after convolution; d=f(z1)∈R m×w , and f is the corresponding one-dimensional deconvolution layer; Hidden variable z2∈R m′×w , is the embedded representation compressed along the feature dimension on the reconstructed d, and m' is the feature dimension after compression; Latent variables represent the potential characteristics or representations of the input data; Hidden variable z1 of time correlation; Latent variable z2 of spatial correlation; After obtaining the latent variable z1 of time correlation and the latent variable z2 of space correlation, the difference or loss between the original data and the reconstructed data is calculated through the reconstruction error function (L) to achieve the minimum or optimize to the minimum; in, represents the expected logarithmic difference between the probability of data in the original distribution and the approximate distribution, p θ represents the probability distribution of the Kullback-Leibler divergence, e represents the reconstructed input required by the generative network (pnet), and q φ represents the upper approximate distribution of the Kullback-Leibler divergence; The way to model normal patterns in telemetry data is to: Based on the self-attention mechanism of the GRU neural network, the model constructed for the normal mode of telemetry data is expressed as: p θ (x,z2,z1)=p θ (x|z2,z1)p θ (z2|z1)p θ (z1); By jointly training random latent variables, z1 and z2 learn time embedding or spatial feature embedding; the hierarchical structure of the layered variational autoencoder is used to transform the original input x into a low-dimensional representation z (z2|z1 hierarchy); Perform a variational posteriors on the model: q φ (z2,z1|x)=q φ (z2|z1)q φ (z1|x); Get the attention weights at different positions in the input telemetry data sequence: a t =softmax(w2 tanh(w1c t +b1)); where w1, w2, and b1 represent the weights and biases of the corresponding fully connected layers.

2. The satellite multi-dimensional telemetry sequence anomaly detection model construction method according to claim 1 is characterized in that: The temporally correlated latent variables are mapped via a convolutional network.

3. The satellite multi-dimensional telemetry sequence anomaly detection model construction method according to claim 1 is characterized in that: The step of reconstructing the input information is obtained through the deconvolution network.

4. The satellite multi-dimensional telemetry sequence anomaly detection model construction method according to claim 3 is characterized in that: The time-correlated latent variables are mapped back to the original data space through a deconvolution network to obtain reconstructed input information.

5. The satellite multi-dimensional telemetry sequence anomaly detection model construction method according to claim 1, characterized in that: Through GRU and attention mechanism, the latent variables of spatial correlation are obtained.

6. A device for constructing a satellite multi-dimensional telemetry sequence anomaly detection model, characterized in that: The device comprises: Module for collecting telemetry data; A module for mapping the data into time-dependent latent variables; According to the latent variables of the time correlation, a module for reconstructing input data is obtained; A module for obtaining latent variables of spatial correlation according to the reconstructed input data; A module for obtaining the difference between the telemetry data and the reconstructed data according to the latent variables of the spatial correlation; For the data X∈R M×N , each row x i is called a feature, and each column x t is called an observation; Hidden variable z1=g(x)∈R m×w′ , g is a number of one-dimensional convolutions, m is the number of features, and w' is the length of the window after convolution; d=f(z1)∈R m×w , and f is the corresponding one-dimensional deconvolution layer; Hidden variable z2∈R m′×w , is the embedded representation compressed along the feature dimension on the reconstructed d, and m' is the feature dimension after compression; Latent variables represent the potential characteristics or representations of the input data; Hidden variable z1 of time correlation; Latent variable z2 of spatial correlation; After obtaining the latent variable z1 of time correlation and the latent variable z2 of space correlation, the difference or loss between the original data and the reconstructed data is calculated through the reconstruction error function (L) to achieve the minimum or optimize to the minimum; in, represents the expected logarithmic difference between the probability of data in the original distribution and the approximate distribution, p θ represents the probability distribution of the Kullback-Leibler divergence, e represents the reconstructed input required by the generative network (pnet), and q φ represents the upper approximate distribution of the Kullback-Leibler divergence; The way to model normal patterns in telemetry data is to: Based on the self-attention mechanism of the GRU neural network, the model constructed for the normal mode of telemetry data is expressed as: p θ (x,z2,z1)=p θ (x|z2,z1)p θ (z2|z1)p θ (z1); By jointly training random latent variables, z1 and z2 learn time embedding or spatial feature embedding; the hierarchical structure of the layered variational autoencoder is used to transform the original input x into a low-dimensional representation z (z2|z1 hierarchy); Perform a variational posteriors on the model: q φ (z2,z1|x)=q φ (z2|z1)q φ (z1|x); Get the attention weights at different positions in the input telemetry data sequence: a t =softmax(w2 tanh(w1c t +b1)); where w1, w2, and b1 represent the weights and biases of the corresponding fully connected layers.

7. A satellite multi-dimensional telemetry sequence anomaly detection method, characterized in that: The method comprises: Steps for collecting the measured telemetry data sequence; The method according to claim 1, comprising the step of obtaining a difference between the telemetry data and the reconstructed data of the sequence.

8. Satellite multi-dimensional telemetry sequence anomaly detection device, characterized in that: The device comprises: A module for collecting the measured telemetry data sequence; The apparatus according to claim 6, comprising means for obtaining a difference between the telemetry data and the reconstructed data of the sequence.

9. A computer storage medium for storing a computer program, characterized in that: When the program is read by a computer, the computer executes the method of claim 1 to 5 or 7.

10. A computer, comprising a processor and a storage medium, characterized in that: When the processor reads the computer program stored in the storage medium, the computer executes the method of claim 1-5 or 7.

Images