Multi-dimensional telemetering sequence anomaly detection method and device based on improved layered variational automatic encoder

Abstract

The invention discloses a multi-dimensional telemetering sequence anomaly detection method and device based on an improved layered variational automatic encoder, and belongs to the field of data processing and detection. The invention provides a multi-dimensional telemetry sequence anomaly detection method for solving the technical problems that an existing multi-dimensional telemetry sequence anomaly detection algorithm is insufficient in modeling of time-space association of telemetry data and labeling accuracy of an anomaly labeling strategy. The method comprises the following steps: collecting telemetry data; mapping the data into latent variables of time correlation; obtaining reconstruction input data according to the latent variable of the time correlation; obtaining a latent variable of spatial correlation according to the reconstructed input data; obtaining a reconstruction error between the telemetering data and the reconstruction data according to the latent variable of the temporal correlation and the latent variable of the spatial correlation; calculating an abnormal score of the time window; and fusing the reconstruction error and the anomaly score of the time window to carry out anomaly marking and the like. The method can be applied to the field of spacecraft telemetry sequence anomaly detection.

Description

Technical Field

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

[0002] For multi-dimensional telemetry sequences, many traditional telemetry data anomaly detection methods, including those based on statistical models and time series decomposition, have been used for telemetry sequence anomalies. For example, the Local Outlier Factor algorithm (LOF) based on statistical eigenvalues, the Autoregressive Moving Average prediction model (ARMA), the Autoregressive Integrated Moving Average model (ARIMA), etc. However, with the improvement of the on-orbit monitoring requirements of spacecraft, telemetry data presents characteristics such as large data volume, high dimension, and complex relationships. Traditional methods have limited adaptability to non-linear relationships when dealing with high-dimensional data and are difficult to effectively capture the complex correlations of the system, resulting in the inability to detect complex potential anomaly patterns in the data.

[0003] To overcome the limitations of traditional methods, in recent years, deep learning methods have received wide attention in the field of anomaly detection. Since obtaining a large amount of labeled data is usually expensive and time-consuming, many methods adopt unsupervised learning methods to learn the normal patterns of telemetry sequences and measure the degree of anomalies based on prediction and reconstruction expectation errors. These methods can automatically learn higher-level abstract representations from the data by constructing a series of popular architectures, such as Temporal Convolutional Networks (TCN), Long Short-Term Memory Networks (LSTMs), and Transformers, effectively improving the performance of time series anomaly detection.

[0004] However, there are some drawbacks when directly applying existing prediction or reconstruction-based methods to telemetry sequence anomaly detection. First, as a complex system, the changes in spacecraft telemetry data may be triggered or regulated by specific telemetry commands, and there is usually a delay in the interaction between commands and telemetry sequence data. Therefore, there are complex interaction correlations and spatio-temporal dependencies between telemetry data and control commands. However, existing prediction or reconstruction-based methods either use simple deterministic methods, resulting in the model being difficult to capture complex patterns and non-linear relationships when modeling multi-variable telemetry sequences, so the performance is weak in modeling spatial correlations. Or they regard telemetry sequences as independent and identically distributed samples or lack the embedded representation of telemetry sequences in the time series dimension, resulting in the inability to effectively capture the evolution and pattern changes of data in time, and the performance is weak in modeling temporal correlations. In addition, while the above methods focus on improving model performance, they often ignore the problem of insufficient labeling accuracy of multi-telemetry parameter anomaly labeling strategies. The anomaly patterns in telemetry data are diverse and have different durations. Methods that rely only on labels or quantitative errors for anomaly labeling can only label some anomaly samples and anomaly patterns, and there is a risk of missed detection.

[0005] An et al. first applied VAE to anomaly detection. VAE combines the ideas of deep learning and probabilistic modeling and can effectively learn the latent representation of data. Existing research has shown that in the task of time series anomaly detection, anomaly detection methods based on VAE have achieved certain results. However, the samples generated by VAE may have a certain degree of ambiguity. In addition, VAE may have the problem of information loss when dealing with long time series data. As an extension of the multi-layer VAE, HVAE can learn the hierarchical structure of data and, through a multi-level generation process, enable the model to better capture the complex features and structures of data.

[0006] In summary, for the problem of telemetry data anomaly detection, when designing a model, it is necessary to comprehensively consider time dependence, spatial correlation, and the interaction between commands and telemetry sequences. Combining different deep learning techniques, such as LSTM, Transformer, VAE, and graph attention mechanisms, more powerful, accurate, and robust telemetry anomaly detection models can be constructed. The development of these models will help improve the reliability and safety of satellite operation and promote the performance optimization of spacecraft and its related systems.

[0007] However, existing multi-dimensional time series anomaly detection algorithms face the problems of complex spatio-temporal associations between telemetry data and control commands and insufficient marking accuracy of anomaly marking strategies. Summary of the Invention

[0008] Aiming at the problems that existing multi-dimensional time series anomaly detection algorithms are limited by the complex spatio-temporal associations between telemetry data and control commands and the insufficient marking accuracy of anomaly marking strategies, the technical solution provided by the present invention is as follows:

[0009] A multi-dimensional telemetry sequence anomaly detection method (Uni-AE) based on an improved hierarchical variational autoencoder (HVAE), the method comprising:

[0010] Steps of collecting telemetry data;

[0011] Steps of mapping the data into latent variables with time correlation;

[0012] Steps of obtaining reconstructed input data according to the latent variables with time correlation;

[0013] Steps of obtaining latent variables with spatial correlation according to the reconstructed input data;

[0014] Steps of obtaining the reconstruction error between the telemetry data and the reconstructed data according to the latent variables with time correlation and the latent variables with spatial correlation.

[0015] Further, a preferred embodiment is provided, in which the latent variable of the temporal correlation is mapped by a one-dimensional convolutional network.

[0016] Further, a preferred embodiment is provided, which includes the step of obtaining the reconstructed input information through a one-dimensional transposed convolutional network.

[0017] Further, a preferred embodiment is provided, in which the latent variable of the temporal correlation is mapped back to the original data space through a transposed convolutional network to obtain the reconstructed input information.

[0018] Further, a preferred embodiment is provided, in which the latent variable of the spatial correlation is obtained through a variant of the recurrent neural network, namely GRU, and the self-attention mechanism.

[0019] Further, a preferred embodiment is provided, in which the difference (reconstruction error) between the telemetry data and the reconstructed data of the sequence is obtained through the latent variable of the temporal correlation and the latent variable of the spatial correlation.

[0020] Further, a preferred embodiment is provided, in which the reconstruction error and the anomaly score of the calculation time window are fused to capture the anomaly change pattern.

[0021] Based on the same inventive concept, the present invention also provides a multi-dimensional telemetry sequence anomaly detection device based on an improved hierarchical variational autoencoder, and the device includes:

[0022] A module for collecting telemetry data;

[0023] A module for mapping the data into a latent variable of the temporal correlation;

[0024] A module for obtaining the reconstructed input data according to the latent variable of the temporal correlation;

[0025] A module for obtaining the latent variable of the spatial correlation according to the reconstructed input data;

[0026] A module for obtaining the reconstruction error between the telemetry data and the reconstructed data according to the latent variable of the temporal correlation and the latent variable of the spatial correlation.

[0027] A module for fusing the reconstruction error and the anomaly score of the calculation time window to capture the anomaly change pattern.

[0028] Based on the same inventive concept, the present invention also provides a computer storage medium for storing a computer program, and when the program is read by a computer, the computer executes the method.

[0029] Based on the same inventive concept, the present invention also provides a computer, including a processor and a storage medium. When the processor reads the computer program stored in the storage medium, the computer executes the method described above.

[0030] Compared with the prior art, the beneficial effects of the technical solution provided by the present invention are as follows:

[0031] The multi-dimensional telemetry sequence anomaly detection method based on improved HVAE provided by the present invention combines a hierarchical variational autoencoder with two random latent variables to model the complex spatio-temporal correlations of multi-dimensional telemetry sequences and capture long-term dependencies.

[0032] The multi-dimensional telemetry sequence anomaly detection method based on improved HVAE provided by the present invention integrates an anomaly marking strategy for multi-source uncertainty fusion to improve the accurate marking and recognition of continuous anomaly events. It enhances the characterization ability of qualitative indicators and continuous features in classical anomaly marking methods for anomaly events.

[0033] The multi-dimensional telemetry sequence anomaly detection method based on improved HVAE provided by the present invention is applicable to the field of spacecraft telemetry sequence anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic structural diagram of the multi-dimensional telemetry sequence anomaly detection method based on improved HVAE provided for Embodiment 1;

[0035] Figure 2 Schematic flow diagram of the multi-dimensional telemetry sequence anomaly detection method based on improved HVAE provided for Embodiment 7;

[0036] Figure 3 Schematic comparison diagram of the actual anomaly sequence and the anomaly detection result mentioned in Embodiment 10;

[0037] Among them, sub- Figure 1 is the actual anomaly sequence, and the rest are the anomaly detection results of the comparison method.

[0038] Figure 4 Schematic diagram of the effect of the method component ablation experiment in the public dataset mentioned in Embodiment 10;

[0039] Figure 5 Schematic basic structure diagram of the spacecraft power system mentioned in Embodiment 11;

[0040] Figure 6 Schematic diagram of the 8-dimensional temperature telemetry parameters related to the spacecraft power system mentioned in Embodiment 11;

[0041] Figure 7 Schematic diagram of the detection result of the 8-dimensional temperature telemetry parameters related to the spacecraft power system mentioned in Embodiment 11. Detailed implementation manners

[0042] To make the advantages and beneficial effects of the technical solution provided by the present invention more clearly reflected, the technical solution provided by the present invention will be further described in detail below with reference to the accompanying drawings:

[0043] Embodiment 1. In combination with Figure 1 This embodiment is described. This embodiment provides a multi-dimensional telemetry sequence anomaly detection method based on an improved HVAE. The method includes:

[0044] Steps of collecting telemetry data;

[0045] Steps of mapping the data into latent variables with time correlation;

[0046] Steps of obtaining reconstructed input data according to the latent variables with time correlation;

[0047] Steps of obtaining latent variables with spatial correlation according to the reconstructed input data;

[0048] Steps of obtaining the difference between the telemetry data and the reconstructed data according to the latent variables with time correlation and the latent variables with spatial correlation.

[0049] Steps of fusing the reconstruction error and the anomaly score of the calculation time window to capture the anomaly change pattern.

[0050] Specifically:

[0051] The telemetry data x is first processed by a one-dimensional convolutional network. The convolutional layer can extract local features of the input data. Through the combination of convolution and activation functions, x is mapped into a latent representation z2.

[0052] The z2 is input into the deconvolutional network. The deconvolutional network maps the hidden representation z2 back to the original data space through reverse convolutional operations to obtain the reconstructed input d. The reconstructed input d is an approximate reconstruction of the input data x.

[0053] The reconstructed input d is passed as an input to a network similar to SRNN. In the SRNN network, techniques such as the Gated Recurrent Unit (GRU) and the Attention Mechanism are used to process sequence data.

[0054] GRU (Gated Recurrent Unit): GRU is a variant of the Recurrent Neural Network (RNN), designed to process sequential data such as text, time series, etc. It controls the flow of information through gating mechanisms, including the update gate and the reset gate, to better capture long-term dependencies in the sequence.

[0055] Attention Mechanism: It is mainly used to model the relationships between different positions in a sequence.

[0056] By calculating the weights of each position in the sequence, it adaptively selects and weights the information of different positions, combining the reconstructed input d and the previous hidden state, so as to focus on the information with higher importance in the SRNN network.

[0057] Processing to obtain z2: The SRNN network combines GRU and the attention mechanism through calculations in multiple time steps, gradually processes the reconstructed input d and outputs the hidden representation z2. In this way, z2 can be used for subsequent detection tasks.

[0058] Among them,

[0059] 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 value.

[0060] The latent variable z2 = g(x) ∈ m×w′ , where g is several one-dimensional convolutions, m is the number of features, and w' is the length of the window after convolution.

[0061] d = g(z2) ∈ R m×w , and f is the corresponding one-dimensional transposed convolution layer.

[0062] The latent variable z1 ∈ R m′×w is an embedded representation obtained by compressing along the feature dimension on the reconstructed d, and m' is the compressed feature dimension.

[0063] The latent variable represents the latent features or representations of the input data.

[0064] The latent variable z2 with temporal correlation;

[0065] The latent variable z1 with spatial correlation;

[0066] After obtaining the latent variable z1 with temporal correlation and the latent variable z2 with spatial correlation, the reconstruction error function (L) is used to calculate the difference or loss between the original data and the reconstructed data, so as to minimize it or optimize it to the minimum.

[0067]

[0068] Among them, represents the expectation under the variational posterior q φ and is used to calculate the measure of the importance of the reconstructed data e with respect to the observed data x. p θ represents the generative distribution, and e represents the reconstructed input, which is used to assist in the calculation of the generative distribution p θ . q φ represents the variational posterior.

[0069] The model aims to achieve better performance through training. By using the reconstruction error function (formula L), the difference or loss between the original data and the reconstructed data is calculated to be minimized or optimized to the minimum. For example, there are many sets of weights, and I'm not sure which set is suitable for my model. Through training, the model finds the suitable set to make the loss of formula L smaller, that is, the performance of the model is better.

[0070] In the anomaly detection problem, we hope to identify the anomalous data points that do not conform to the normal pattern. A common approach is to use a reconstruction model, which attempts to generate approximate copies or reconstructions from the original data and measures the degree of anomaly by comparing the difference between the original data and the reconstructed data.

[0071] 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 high reconstruction probability because the model can reconstruct these data points well. For anomalous data points, they may be quite different from the normal pattern, resulting in the model having difficulty accurately reconstructing them, so they have a low reconstruction probability.

[0072] The generative network (pnet in the figure) is part of the HVAE and is responsible for generating samples from the latent space. The generative network receives a random vector from the latent variable space (usually following the standard normal distribution) as input and, through a series of mapping and transformation operations, converts it into a generated sample with similar characteristics to the original data. The goal of the generative network is to learn the distribution of the data so as to be able to generate new samples similar to the training data.

[0073] The variational posterior network (qnet in the figure) is also an important part of the HVAE and 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, through a series of mapping and transformation operations, converts them into the distribution parameters of the latent variables. The goal of the variational posterior network is to learn the distribution of the data in the latent space so as to be able to perform inference and reconstruction of the latent variables for the given input data.

[0074] The generative network and the variational posterior network are interrelated. In HVAE, the generative network and the variational posterior network are jointly trained 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 samples 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 collaboratively, enabling HVAE to learn the ability to generate samples and infer latent variables simultaneously.

[0075] Model architecture part:

[0076] The telemetry sequences of spacecraft exhibit complex interdependencies, including temporal correlations (such as periodicity) and variable correlations (linear or non-linear relationships within the variables of each time period entity). A hierarchical variational autoencoder with two stochastic latent variables can model the normal modes of telemetry sequences. On this basis, we integrate a self-attention mechanism based on GRU neural network, enabling 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 as Figure 1 shown. Specifically, the generative model (pnet) can be expressed as:

[0077] p θ (x, z1, z2) = p θ (x|z1, z2)p θ (z1|z2)p θ (z1);

[0078] By jointly training the stochastic latent variables, z2 and z1 learn temporal embeddings or spatial feature embeddings. The hierarchical structure of the hierarchical variational autoencoder is used to transform the original input x into a low-dimensional representation z (the z1|z2 hierarchy).

[0079] The purpose of this hierarchy is to enable z1 to combine z2 to capture the learned temporal information, rather than learning spatial feature embeddings independently. Specifically, for the temporal embedding layer, z2 = g(x) ∈ R m×w′ , where 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 = g(z2) ∈ R m×w and f is the corresponding one-dimensional transposed convolution layer. The spatial feature embedding z1 ∈ R m′×w is the embedding representation compressed along the feature dimension on the reconstructed d, and m′ is the compressed feature dimension. The variational autoencoder VAE is used to pre-train d to ensure its initial reconstructability in the initial stage of model training. As Figure 1 shown, the variational posterior (qnet) can be expressed as:

[0080] q φ (z1, z2|x) = q φ (z1|z2)q φ (z2|x);

[0081] For each time instant t, the input values d t and a t+1 (the deterministic state derived by the GRU at time instant t + 1) are passed to a reverse recurrent GRU cell:

[0082]

[0083] Finally, in addition, a self-attention mechanism is utilized in the GRU network to better handle the long-term dependencies and context information in the input sequence:

[0084] A = softmax(w2tanh(w1H T + b1))·H;

[0085] where w1, w2, and b1 represent the weights and biases of the corresponding fully connected layers. H is the hidden state matrix of the input sequence. In particular, the parameters of the one-dimensional deconvolution (for deriving d, e) layer are shared between the generation network and the variational posterior network, so that the "reconstructed input" and the obtained temporal information can be shared, thereby improving the training effect. Finally, a planar NF transformation is applied in qnet to capture the latent non-Gaussian posterior distribution of the data by introducing a series of invertible planar transformations.

[0086] Embodiment 2. This embodiment further limits the method for anomaly detection of multi-dimensional telemetry sequences based on the improved HVAE provided in Embodiment 1, and maps the latent variable of the time correlation through a convolutional network.

[0087] Embodiment 3. This embodiment further limits the method for anomaly detection of multi-dimensional telemetry sequences based on the improved HVAE provided in Embodiment 1, and the step of obtaining the reconstructed input information through a deconvolution network.

[0088] Embodiment 4. This embodiment further limits the method for anomaly detection of multi-dimensional telemetry sequences based on the improved HVAE provided in Embodiment 3, and maps the latent variable of the time correlation back to the original data space through a deconvolution network to obtain the reconstructed input information.

[0089] Embodiment 5. This embodiment further limits the method for anomaly detection of multi-dimensional telemetry sequences based on the improved HVAE provided in Embodiment 1, and obtains the latent variable of the spatial correlation through GRU and the attention mechanism.

[0090] Embodiment 6. This embodiment provides a multi-dimensional telemetry sequence anomaly detection device based on an improved HVAE. The device includes:

[0091] A module for collecting telemetry data;

[0092] A module for mapping the data into latent variables with time correlation;

[0093] A module for obtaining the reconstructed input data based on the latent variables with time correlation;

[0094] A module for obtaining the latent variables with spatial correlation based on the reconstructed input data;

[0095] A module for obtaining the reconstruction error between the telemetry data and the reconstructed data based on the latent variables with time correlation and the latent variables with spatial correlation.

[0096] A module for fusing the reconstruction error and the anomaly score of the calculation time window to capture the anomaly change pattern.

[0097] Embodiment 7. In combination with Figure 2 Describe this embodiment. According to the method provided in Embodiment 1, obtain the reconstruction error between the telemetry data and the reconstructed data of the sequence, and fuse the reconstruction error and the anomaly score of the calculation time window to capture the anomaly change pattern.

[0098] Specifically:

[0099] Steps of training and inference:

[0100] Optimize the evidence lower bound (ELBO) through the SGVB algorithm, where D KL The divergence is the Kullback-Leibler divergence. The mathematical formula is as follows:

[0101]

[0102] After introducing the reconstructed inputs d and e, the model optimization objective becomes the following form:

[0103]

[0104] During online detection, our goal is to detect whether the timestamp x t is an anomaly through a directional sliding window of length W and calculate the reconstruction probability This is widely used in anomaly detection. The reconstruction probability of the last data x t in this window is used as the anomaly score at time t.

[0105] Anomaly detection judgment process:

[0106] The trained model is used to determine anomalies in the observed values at time t. A sliding window is selected, that is, t and the previous w - 1 consecutive observed values are used as inputs to detect anomalies at time t. The last data x t in the sliding window is used as the anomaly score, and the anomaly score of x t is denoted as S t . In practice, anomalies are rarely independent. More commonly, there are continuous anomaly segments. While marking anomalies, the number of marked anomalies in the corresponding window is recorded, and the anomaly score at the corresponding time t is obtained by optimizing the reconstruction probability of the model. The anomaly degree is comprehensively evaluated by combining the reconstruction error and the number of historical window anomalies. Formally, if S t is higher than the predefined threshold, the observed value x t is marked as an anomaly; otherwise, x t is normal. The mathematical formula is as follows:

[0107]

[0108] The average value of the reconstruction probability scores of the last k data in each sliding window is used to evaluate anomalies. I i is an indicator function indicating whether the i-th observed value in the window is marked as an anomaly. Figure 3 Describes an example of the judgment of anomaly detection. As the spacecraft telemetry data arrives one after another, the anomaly scores corresponding to the data are obtained through the reconstruction error, and are compared with the threshold. Among them, the orange line is the calculated threshold, the blue line represents the anomaly scores corresponding to the data, and the red area is the determined anomaly segment.

[0109] Embodiment 8. This embodiment provides a computer storage medium for storing a computer program. When the program is read by a computer, the computer executes the methods provided in Embodiments 1 to 7.

[0110] Embodiment 9. This embodiment provides a computer, including a processor and a storage medium. When the processor reads the computer program stored in the storage medium, the computer executes the methods provided in Embodiments 1 to 7.

[0111] Embodiment 10. Combining Figure 3 with Figure 4 to illustrate this embodiment. This embodiment is an example for providing comparative verification and ablation verification for the method provided in Embodiment 7. Specifically:

[0112] Table 1 Parameter Configuration

[0113]

[0114] 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 iteration continues for 20 epochs. RELU is used as the activation function for layers other than 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 adopted during training to prevent the model from overfitting.

[0115] Performance of the comparison methods in Table 2

[0116]

[0117] As shown in Table 2, it can be seen that the designed model achieves the best F1 scores of 0.9413 and 0.9473 on the SMAP dataset and the MSL dataset. Specifically, the following analysis can be made.

[0118] Specifically, compared with Uni-AE, UniAE-T removes the latent variable z1 and only uses the latent variable z2 to capture the learned temporal information of the telemetry data. On the contrary, UniAE-S removes the latent variable z2 and only uses the latent variable z1 to capture the learned spatial correlation of the telemetry data. UniAE-p adopts independent latent variables z2 and z1 in the variational network, that is, q φ (z1, z2|x) = q φ (z1|x)q φ (z2|x). And self-attention is removed before variational inference, and only the GRU model is used to learn the dynamic features in the time series and output the hidden state sequence. UniAE-x does not apply the anomaly marking strategy of multi-source uncertainty fusion, which makes the anomaly score of the observation point only depend on the reconstruction probability of the corresponding point.

[0119] The proposed Uni-AE uses a hierarchical variational autoencoder with two latent variables to learn the spatio-temporal correlation of multi-dimensional telemetry sequences. The latent state sequence is input into the attention mechanism in the GRU, and relevant information in the input sequence is adaptively selected and weighted according to the attention weights. The experimental results show the importance of each component in Uni-AE for the model performance, further verifying the effectiveness of the method we proposed.

[0120] Compared with other baselines, the proposed method shows superior performance, with a difference of 0.04 - 0.2308 in the SMAP dataset and a difference of 0.0365 - 0.3833 in the MSL dataset. A hierarchical variational autoencoder that combines two random latent variables is used to model the complex spatio-temporal associations of multi-dimensional telemetry sequences and capture long-term dependencies. An anomaly labeling strategy based on multi-source uncertainty fusion is combined to improve the accurate labeling and recognition of continuous anomaly events. The ability of qualitative indicators and continuous features in classical anomaly labeling methods to characterize anomaly events is enhanced. This makes it sufficiently attractive in dealing with anomaly detection of multi-dimensional telemetry data.

[0121] Embodiment XI. Combination Figure 5 , 6 , 7 describes this embodiment. This embodiment is an example for case verification of the method provided in Embodiment VII. Specifically:

[0122] Table 3 Performance of Each Method

[0123]

[0124] Figure 5 Describe the basic structure of the spacecraft power system. Using an 8-dimensional temperature telemetry sequence as the input, which are the satellite battery wall temperature and the temperatures of two groups of batteries respectively, and its curve is as Figure 6 shown. Abnormal patterns appeared in the temperature data of multiple telemetry parameters within the time index range of 6902 to 8467. The performance of all comparison methods on the satellite power system data is shown in Table 3. It can be seen that the proposed method shows higher Precision and Recall, indicating that our method has a lower false alarm rate and missed alarm rate when detecting anomalies. Figure 7 Show the application effect of the anomaly detection method on the power system dataset, including normal values, anomaly values, and thresholds marked with different colors.

[0125] The above further describes the technical solutions provided by the present invention through several specific embodiments to highlight the advantages and beneficial effects of the technical solutions provided by the present invention. However, the above-mentioned several specific embodiments are not used as limitations to the present invention. Any reasonable modifications and improvements to the present invention, reasonable combinations of embodiments, and equivalent replacements within the spirit and principle of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-dimensional telemetry sequence anomaly detection method based on an improved hierarchical variational autoencoder, 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 a latent variable of spatial correlation according to the reconstructed input data; The step of obtaining a reconstruction error between the telemetry data and the reconstructed data according to the latent variable of the time correlation and the latent variable of the space correlation. The step of fusing the reconstruction error and the anomaly score of the calculation time window to capture the abnormal change pattern.

2. The multidimensional telemetry sequence anomaly detection method based on improved hierarchical variational autoencoder HVAE according to claim 1 is characterized in that: The temporally correlated latent variables are mapped via a convolutional network.

3. The multidimensional telemetry sequence anomaly detection method based on improved hierarchical variational autoencoder HVAE according to claim 1 is characterized in that: The step of reconstructing the input information is obtained through the deconvolution network.

4. The multidimensional telemetry sequence anomaly detection method based on improved hierarchical variational autoencoder HVAE according to claim 3 is characterized in that: The time-dependent latent variables are mapped back to the original data space through a deconvolution network to obtain reconstructed input information.

5. The multi-dimensional telemetry sequence anomaly detection method based on improved hierarchical variational autoencoder HVAE according to claim 1 is characterized in that: Through GRU and self-attention mechanism, latent variables of spatial correlation are obtained.

6. A multi-dimensional telemetry sequence anomaly detection device based on an improved hierarchical variational autoencoder HVAE, 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 a reconstruction error between the telemetry data and the reconstructed data according to the latent variables of the time correlation and the latent variables of the space correlation; A module that captures abnormal change patterns based on the reconstruction error and the anomaly score of the calculation time window.

7. 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 claims 1 to 5.

8. 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 claims 1 to 5.

Images