Satellite on-orbit telemetry data anomaly detection method and system

Abstract

This invention discloses a method and system for anomaly detection in satellite on-orbit telemetry data. The method involves acquiring the on-orbit telemetry time-series data of the satellite to be tested; decomposing the data into multiple approximate and detail sequences using the time-frequency characteristics of multi-scale wavelet transform; assigning weights to each approximate and detail sequence, calculating attention weights for each sequence by calculating similarity scores; weighted summing of the hidden states of neurons in the coding layer and their corresponding attention weights within each approximate and detail sequence to calculate a content vector, which is then combined with the current hidden state of neurons in the decoding layer, and so on, to obtain a reconstructed sequence of the on-orbit telemetry time-series data; and outputting anomaly data by comparing the difference between the reconstructed sequence and the on-orbit telemetry time-series data of the satellite to be tested. This method effectively processes the spatiotemporal information of satellite telemetry data, improving the accuracy and robustness of anomaly detection.

Description

Technical Field

[0001] This invention relates to the fields of satellite telemetry and artificial intelligence, and more specifically to a method and system for detecting anomalies in satellite telemetry data based on an autoencoder and a long short-term memory network. Background Technology

[0002] Satellite telemetry data is data transmitted to ground receiving stations by the satellite's internal sensor monitoring system during its on-orbit operation. This data includes various operational parameters such as power system status, attitude control, thermal control, and communication system status. Ground personnel analyze this data to monitor the satellite's operational status, identify potential problems, assess subsystem performance, and make adjustments and maintenance. Therefore, real-time monitoring and detection of abnormal satellite operating conditions are crucial for the timely discovery and location of satellite faults.

[0003] Satellite systems consist of numerous subsystems, payloads, and component units, each potentially involving dozens or even hundreds of telemetry parameters. With increasing computing and storage capabilities, the scale of telemetry data from thousands of sensors is constantly growing. As satellite components and devices age, their performance and characteristics change, resulting in long-term dependencies in telemetry data features. Furthermore, complex correlations exist between different telemetry parameters, not only within components but also between different parts of the system. Overall, this telemetry data is essentially a time series, exhibiting complex characteristics such as high dimensionality, nonlinearity, correlation, and multimodality. Therefore, anomaly detection in satellite telemetry data faces challenges such as large data volume, high dimensionality, complex temporal relationships, and strong correlations between parameters. How to improve the accuracy of satellite anomaly detection while ensuring real-time performance is a critical issue that urgently needs to be addressed in current on-orbit satellite real-time monitoring.

[0004] With the continuous development of satellite technology and the expanding application of artificial intelligence, experts have proposed using intelligent methods, leveraging years of accumulated engineering knowledge and powerful AI capabilities, combined with satellite telemetry time-series data, to conduct satellite anomaly detection, thus ensuring the stable and healthy operation of satellites. By training models with AI algorithms, these models can identify anomaly patterns, thereby quickly and accurately detecting abnormal data, significantly improving detection efficiency and accuracy. Furthermore, by utilizing big data and cloud computing technologies, telemetry data can be processed and analyzed in real time, providing immediate anomaly detection and more timely and accurate support for on-orbit satellite management. It is foreseeable that AI-based on-orbit satellite telemetry data anomaly detection and assessment will play a crucial role in subsequent on-orbit satellite management, contributing significantly to supporting the stable and safe operation of satellites. Therefore, researching satellite telemetry data anomaly detection algorithms is imperative for effectively managing the on-orbit operational status of satellites and promptly handling anomalies.

[0005] Because satellite telemetry data is essentially a time series and possesses characteristics such as high dimensionality, nonlinearity, correlation, and multimodality, satellite anomaly detection faces challenges such as processing large volumes of telemetry data in parallel, handling high-dimensional telemetry parameters, and navigating complex temporal relationships and correlations between telemetry data. Therefore, higher demands are placed on the accuracy of satellite telemetry data anomaly detection algorithms.

[0006] Currently, although the traditional LSTM-AE model can capture long-term dependencies in sequences to some extent, the limitation of the length of the intermediate hidden states of LSTM, i.e., the intermediate codes, makes it difficult for the LSTM-AE model to retain all the information in particularly long-term sequences. It cannot effectively capture and utilize long-distance dependencies, which ultimately leads to a decrease in the accuracy of the model when reconstructing sequences with long-term dependencies. Summary of the Invention

[0007] To address the problems existing in the above-mentioned fields, this invention proposes a method and system for detecting anomalies in satellite on-orbit telemetry data. This method can solve the technical problem that although the traditional LSTM-AE model can capture long-term dependencies in the sequence to a certain extent, the limitation of the length of the intermediate hidden state of LSTM, i.e., the intermediate code, makes it difficult for the LSTM-AE model to retain all the information in a particularly long-term sequence. As a result, it cannot effectively capture and utilize long-distance dependencies, which ultimately leads to a decrease in the accuracy of the model when reconstructing sequences with long-term dependencies.

[0008] To address the aforementioned technical problems, this invention discloses a method for detecting anomalies in satellite on-orbit telemetry data, comprising the following steps:

[0009] Acquire on-orbit telemetry time series data of the satellite to be tested;

[0010] Multi-scale wavelet transform is introduced, and the time-frequency characteristics of multi-scale wavelet transform are used to decompose the satellite on-orbit telemetry time series data into multiple approximate sequences and detail sequences. Weights are assigned to each approximate sequence and detail sequence, and attention weights for each approximate sequence and detail sequence are calculated by calculating similarity scores.

[0011] The hidden states of each neuron in the coding layer and their corresponding attention weights within each approximate sequence and detail sequence are weighted and summed to calculate the content vector; the content vector is combined with the current hidden state of the neuron in the decoding layer, and so on, to obtain the reconstructed sequence of satellite on-orbit telemetry time series data;

[0012] By comparing the difference between the reconstructed sequence and the on-orbit telemetry time series data of the satellite to be tested, abnormal data is output when the difference is greater than a set threshold.

[0013] Preferably, the step of decomposing satellite on-orbit telemetry time-series data into multiple approximate sequences and detail sequences using the time-frequency characteristics of multi-scale wavelet transform includes the following steps:

[0014] A three-level Daubechies wavelet transform is used to decompose the satellite on-orbit telemetry time series data structure, which is decomposed into the first-level approximate sequence AS1 and the detail sequence DS1. The AS1 component is further decomposed into the second-level approximate sequence AS2 and the detail sequence DS2, and so on to complete the three-level Daubechies wavelet decomposition. The three-level approximate sequence AS3 and the four sub-sequences of the three-level detail sequences DS1, DS2 and DS3 are retained.

[0015] Preferably, it also includes calculating an attention score, comprising the following steps:

[0016] By focusing on the attention of the query and the key, that is, given a query, the relevance between the query and the key is calculated;

[0017] Determine the most suitable value based on the relevance between the query and the key;

[0018] Attention scores are calculated by evaluating the similarity between the query at each position in the input sequence and the key at all positions.

[0019] Preferably, calculating the attention weights for each approximate sequence and detail sequence includes the following steps:

[0020] By normalizing the attention score, it is transformed into a probability distribution in which the sum of the weights of all elements is 1, and this distribution is used as the weight coefficient of the corresponding position Value.

[0021] The result of the attention mechanism is obtained by weighting and summing the normalized weight coefficients with the Value.

[0022] The encoder and decoder are integrated with LSTM recurrent units as the basic structure, and weights are assigned to each approximate sequence and detail sequence through the attention mechanism in the Transformer layer.

[0023] Before model decoding, the Transformer layer obtains the hidden states of all sequences in the encoding stage through the calculation results of the attention mechanism; it calculates a similarity score based on the hidden states of all sequences in the encoding stage and the hidden states of the decoding layer; and it calculates the weights 'a' of the attention mechanism's calculation results using the softmax function. ij ;

[0024] Among them, a ij The element to be encoded represents the correlation between other elements and the element with higher correlation; the element with higher correlation has a weight value 'a'. ij Larger.

[0025] Preferably, obtaining the reconstructed sequence of satellite on-orbit telemetry time series data includes the following steps:

[0026] By analyzing the hidden states of each neuron in the coding layer within each approximate sequence and detail sequence. The weight a of the corresponding attention mechanism calculation result ij Calculate the content vector;

[0027] Content vector With the current hidden state of the neuron By combining these, we obtain the reconstruction vector of the current neuron's hidden state:

[0028]

[0029] Among them, W c It is a learnable weight matrix. This indicates that the content vector and the decoder's hidden state are concatenated;

[0030] Will The reconstructed sequence is obtained after passing through a linear layer. Complete the model reconstruction.

[0031] Preferably, the output of abnormal data includes the following steps:

[0032] The acquired on-orbit telemetry time series data of the satellite to be tested will be used as the real data sequence;

[0033] By comparing the difference between the reconstructed sequence and the real data sequence, an outlier is identified when the difference exceeds a set threshold.

[0034] Preferably, the method further includes inverse normalization of the reconstructed sequences of each output approximate sequence and detail sequence to obtain multi-scale reconstruction results, wherein the inverse normalization formula is:

[0035] X original =X normalized *std+mean

[0036] Among them, X original It is the value after denormalization, X normalized These are the normalized values, where mean and std are the mean and standard deviation of the original data, respectively.

[0037] Preferably, the method also includes preprocessing the acquired on-orbit telemetry data of the satellite to be tested, including cleaning and integration.

[0038] Preferably, it also includes a satellite on-orbit telemetry data anomaly detection system, comprising:

[0039] The data acquisition unit is used to acquire on-orbit telemetry time series data of the satellite to be tested;

[0040] The time series data decomposition unit is used to introduce multi-scale wavelet transform and use the time-frequency characteristics of multi-scale wavelet transform to decompose the satellite on-orbit telemetry time series data into multiple approximate sequences and detail sequences.

[0041] The data reconstruction unit is used to assign weights to each approximate sequence and detail sequence. By calculating the similarity score, the attention weight of each approximate sequence and detail sequence is calculated. The hidden states of each neuron in the coding layer and their corresponding attention weights are weighted and summed to calculate the content vector. The content vector is combined with the current hidden state of the neurons in the decoding layer, and so on, to obtain the reconstructed sequence of satellite on-orbit telemetry time series data.

[0042] The anomaly detection unit is used to compare the difference between the reconstructed sequence and the real data sequence, and output abnormal data when the difference is greater than a set threshold.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The proposed anomaly detection method overcomes the limitations of LSTM's intermediate hidden states, which prevent LSTM-AE models from retaining all information in very long-term sequences and effectively capturing and utilizing long-distance dependencies, ultimately leading to decreased accuracy when reconstructing sequences with long-term dependencies. This method introduces multi-scale wavelet transform, utilizing its time-frequency characteristics to decompose satellite telemetry time-series data into multiple scales. This helps the model capture the temporal features of the data more comprehensively and accurately, allowing the model to focus more on capturing important features within the data. Weights are assigned to each approximate sequence and detail sequence. Similarity scores are calculated to weight the input sequence, and a softmax function is used to calculate more precise attention weights. These weights characterize the correlation between other elements and the element to be encoded; elements with higher correlation have larger weights. This improves the rationality of calculating the hidden state of a single neuron in the decoding layer and enhances the strong correlation with the hidden states of neurons in the encoding layer, making it easier to capture certain important features in the sequence. The hidden states of neurons in the encoding layer within each approximate and detail sequence are weighted and summed with their corresponding attention weights to calculate the content vector, thus obtaining the reconstructed sequence. By fully utilizing the recurrent structure to capture long-term information, the global dependency of the Transformer layer and the time-series feature extraction and reconstruction functions of the LSTM autoencoder are combined. This effectively processes the spatiotemporal information of satellite telemetry data. The difference between the reconstructed sequence and the real data sequence is compared with a threshold to ultimately output abnormal data, thereby improving the accuracy and robustness of anomaly detection in satellite telemetry time-series data. Attached Figure Description

[0045] Figure 1 This is a flowchart of the data anomaly detection method of the present invention;

[0046] Figure 2 This is a flowchart of the multi-scale wavelet decomposition method of the present invention;

[0047] Figure 3 This is a schematic diagram of the attention mechanism involved in the present invention;

[0048] Figure 4 This is a diagram of the MS-LSTM-TAE model architecture constructed according to the present invention. Detailed Implementation

[0049] The following will refer to the appendices in the embodiments of the present invention. Figure 1-4 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0050] Example

[0051] like Figure 1 As shown, this embodiment of the invention provides a method for detecting anomalies in satellite on-orbit telemetry data, including the following steps:

[0052] S1: Acquire the on-orbit telemetry time series data of the satellite to be tested, and clean and integrate it.

[0053] S2: Introduce multi-scale wavelet transform and use its time-frequency characteristics to decompose satellite on-orbit telemetry time series data into multiple approximate sequences and detail sequences; assign weights to each approximate sequence and detail sequence, and calculate the attention weight of each approximate sequence and detail sequence by calculating similarity scores.

[0054] S3: Calculate the content vector by summing the hidden states of each neuron in the coding layer and their corresponding attention weights within each approximate sequence and detail sequence; combine the content vector with the current hidden state of the neuron in the decoding layer, and so on, to obtain the reconstructed sequence of satellite on-orbit telemetry time series data.

[0055] S4: By comparing the difference between the reconstructed sequence and the real data sequence, abnormal data is output when the difference exceeds a set threshold.

[0056] In step S2, the satellite's on-orbit telemetry time series data is decomposed into multiple approximate sequences and detail sequences using the time-frequency characteristics of multi-scale wavelet transform, including the following steps:

[0057] A three-level Daubechies wavelet transform is used to decompose the satellite's on-orbit telemetry time-series data structure. The data is decomposed into a first-level approximate sequence AS1 and a detail sequence DS1. The AS1 component is further decomposed into a second-level approximate sequence AS2 and a detail sequence DS2, and so on, completing the three-level Daubechies wavelet decomposition. The third-level approximate sequence AS3 and the four sub-sequences of the three-level detail sequences DS1, DS2, and DS3 are retained. These are then used for training, prediction, and testing of the MS-LSTM-TAE model (the anomaly detection model constructed in this application), and the prediction results are denormalized. Finally, multi-scale wavelet reconstruction is performed to obtain the reconstructed time-domain telemetry data, thereby improving the model's detection accuracy. The processing procedure is as follows: Figure 2 As shown.

[0058] This application constructs an anomaly detection model for satellite on-orbit telemetry data based on LSTM, multi-scale wavelet transform, and an autoencoder-based Transformer, namely the MS-LSTM-TAE model. The model uses an attention mechanism in the Transformer layer to assign weights and calculate attention weights, including the following steps:

[0059] By focusing on the attention of the query and the key, that is, given a query, the relevance between the query and the key is calculated;

[0060] Based on the relevance between the query and the key, the most suitable value is determined, attention weights are allocated to the value, and the final output is generated.

[0061] Specifically, it can be summarized into the following three steps:

[0062] First, an attention score is obtained by calculating the similarity between the query at each position in the input sequence and the key at all positions;

[0063] Second, the attention score is normalized and transformed into a probability distribution in which the sum of the weights of all elements is 1, and used as the weight coefficient of the corresponding position Value.

[0064] Third, the weighted sum of the normalized weight coefficients and the Value is used as the calculation result of the attention mechanism, such as... Figure 3 As shown.

[0065] Before model decoding, the Transformer layer obtains the hidden states of all sequences in the encoding stage through the calculation results of the attention mechanism; it calculates a similarity score based on the hidden states of all sequences in the encoding stage and the hidden states of the decoding layer; and it calculates the weights 'a' of the attention mechanism's calculation results using the softmax function. ij ;

[0066] Among them, a ij The element to be encoded represents the correlation between other elements and the element with higher correlation; the element with higher correlation has a weight value 'a'. ij Larger, making a ij It can improve the rationality of calculating the hidden state htD of a single neuron in the decoding layer and the strong correlation with the hidden states of each neuron in the coding layer.

[0067] Obtaining the reconstructed sequence of satellite on-orbit telemetry time-series data includes the following steps:

[0068] Hidden states of each neuron in the coding layer The weight a of the corresponding attention mechanism calculation result ij Calculate the content vector;

[0069] Content vector With the current hidden state of the neuron By combining these, we obtain the reconstruction vector of the current neuron's hidden state:

[0070]

[0071] Among them, W c It is a learnable weight matrix. This indicates that the content vector and the decoder's hidden state are concatenated;

[0072] Will The reconstructed sequence is obtained after passing through a linear layer. Complete the model reconstruction.

[0073] In step S3, the positional information of the input sequence is embedded through positional encoding in the Transformer layer, enabling the MS-LSTM-TAE model constructed in this application to process sequence data and capture its order information. Local feature representations are extracted independently at each position through a feedforward neural network, helping the model to better understand the input sequence.

[0074] In step S4, abnormal data is output, including the following steps:

[0075] The acquired on-orbit telemetry time series data of the satellite to be tested will be used as the real data sequence;

[0076] By comparing the difference between the reconstructed sequence and the real data sequence, an outlier is identified when the difference exceeds a set threshold.

[0077] like Figure 4 The diagram shows the MS-LSTM-TAE model architecture constructed in this application. This model mainly consists of an encoding layer, a Transformer layer, and a decoding layer. The hidden layer and reconstruction vector calculations during the decoding and encoding stages follow the traditional LSTM computation rules. The hidden state at time t during the model's encoding stage is shown. Input vector at the current time and the hidden state of the previous moment Decide.

[0078] Based on this reconstruction method, this application also proposes a satellite on-orbit telemetry data anomaly detection system, including:

[0079] The data acquisition unit is used to acquire on-orbit telemetry time series data of the satellite to be tested;

[0080] The time series data decomposition unit is used to introduce multi-scale wavelet transform and use the time-frequency characteristics of multi-scale wavelet transform to decompose the satellite on-orbit telemetry time series data into multiple approximate sequences and detail sequences.

[0081] The data reconstruction unit is used to assign weights to each approximate sequence and detail sequence. By calculating the similarity score, the attention weight of each approximate sequence and detail sequence is calculated. The hidden states of each neuron in the coding layer and their corresponding attention weights are weighted and summed to calculate the content vector. The content vector is combined with the current hidden state of the neurons in the decoding layer, and so on, to obtain the reconstructed sequence of satellite on-orbit telemetry time series data.

[0082] The anomaly detection unit is used to compare the difference between the reconstructed sequence and the real data sequence, and output abnormal data when the difference is greater than a set threshold.

[0083] This application, based on traditional LSTM-AE model research, constructs an MS-LSTM-TAE model. By introducing multi-scale wavelet transform, it utilizes the time-frequency characteristics of multi-scale wavelet transform to decompose satellite on-orbit telemetry time-series data into multiple approximate and detail sequences. This multi-scale decomposition of telemetry data helps the model capture the temporal features of the data more comprehensively and accurately, allowing the model to focus more on capturing important features. Simultaneously, it integrates a Transformer based on LSTM recurrent units to weight the input sequence, making it easier to capture certain important features. It also fully utilizes the recurrent structure to capture long-term information, combining the global dependency of the Transformer layer with the time-series feature extraction and reconstruction functions of the LSTM autoencoder. This more effectively processes the spatiotemporal information of satellite telemetry data. By comparing the difference between the reconstructed sequence and the real data sequence, it outputs anomalous data, thereby improving the accuracy and robustness of anomaly detection in satellite telemetry data.

[0084] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0085] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A method for detecting anomalies in satellite on-orbit telemetry data, characterized in that, Includes the following steps: Acquire on-orbit telemetry time series data of the satellite to be tested; Multi-scale wavelet transform is introduced, and the time-frequency characteristics of multi-scale wavelet transform are used to decompose the satellite on-orbit telemetry time series data into multiple approximate sequences and detail sequences. Weights are assigned to each approximate sequence and detail sequence, and attention weights for each approximate sequence and detail sequence are calculated by calculating similarity scores. The hidden states of each neuron in the coding layer and their corresponding attention weights within each approximate sequence and detail sequence are weighted and summed to calculate the content vector; the content vector is combined with the current hidden state of the neuron in the decoding layer, and so on, to obtain the reconstructed sequence of satellite on-orbit telemetry time series data; By comparing the difference between the reconstructed sequence and the on-orbit telemetry time series data of the satellite to be tested, abnormal data is output when the difference is greater than a set threshold. The method of decomposing satellite on-orbit telemetry time-series data into multiple approximate sequences and detail sequences using the time-frequency characteristics of multi-scale wavelet transform includes the following steps: A three-level Daubechies wavelet transform is used to decompose the satellite on-orbit telemetry time series data structure, which is decomposed into the first-level approximate sequence AS1 and the detail sequence DS1. The AS1 component is further decomposed into the second-level approximate sequence AS2 and the detail sequence DS2, and so on to complete the three-level Daubechies wavelet decomposition. The third-level approximate sequence AS3 and the four sub-sequences of the three-level detail sequences DS1, DS2 and DS3 are retained. It also includes calculating an attention score, which involves the following steps: By focusing on the attention of the query and the key, that is, given a query, the relevance between the query and the key is calculated; Determine the most suitable value based on the relevance between the query and the key; The attention score is obtained by calculating the similarity between the query at each position in the input sequence and the key at all positions. The calculation of attention weights for each approximate sequence and detail sequence includes the following steps: By normalizing the attention score, it is transformed into a probability distribution in which the sum of the weights of all elements is 1, and this distribution is used as the weight coefficient of the corresponding position Value. The result of the attention mechanism is obtained by weighting and summing the normalized weight coefficients with the Value. The encoder and decoder are integrated with LSTM recurrent units as the basic structure, and weights are assigned to each approximate sequence and detail sequence through the attention mechanism in the Transformer layer. Before decoding, the Transformer layer obtains the hidden states of all sequences in the encoding stage through the calculation results of the attention mechanism; it calculates a similarity score based on the hidden states of all sequences in the encoding stage and the hidden states of the decoding layer; and it calculates the weights of the attention mechanism's calculation results using the softmax function. a ij ; in, a ij The element to be encoded represents the correlation between other elements and the element with higher correlation; the element with higher correlation has a higher weight. a ij Larger.

2. The method for detecting anomalies in satellite on-orbit telemetry data according to claim 1, characterized in that, The reconstructed sequence of the satellite's on-orbit telemetry time series data includes the following steps: By analyzing the hidden states of each neuron in the coding layer within each approximate sequence and detail sequence. The weights of the corresponding attention mechanism calculation results a ij Calculate the content vector; Content vector With the current hidden state of the neuron By combining these, we obtain the reconstruction vector of the current neuron's hidden state: = fish ( W c [ ; ]) in, W c It is a learnable weight matrix, [ ; The symbol ] indicates that the content vector and the decoder's hidden state are concatenated; Will The reconstructed sequence is obtained after passing through a linear layer. Complete the model reconstruction.

3. The method for detecting anomalies in satellite on-orbit telemetry data according to claim 1, characterized in that, The output of abnormal data includes the following steps: The acquired on-orbit telemetry time series data of the satellite to be tested will be used as the real data sequence; By comparing the difference between the reconstructed sequence and the real data sequence, an outlier is identified when the difference exceeds a set threshold.

4. The method for detecting anomalies in satellite on-orbit telemetry data according to claim 1, characterized in that, It also includes inverse normalization of the reconstructed sequences of each output approximate sequence and detail sequence to obtain multi-scale reconstruction results. The inverse normalization formula is as follows: X original = X normalized * std + mean in, X original It is the value after inverse normalization. X normalized It is the normalized value. mean and std These are the mean and standard deviation of the original data, respectively.

5. The method for detecting anomalies in satellite on-orbit telemetry data according to claim 1, characterized in that, It also includes preprocessing the acquired on-orbit telemetry data of the satellite to be tested, including cleaning and integration.

6. A satellite on-orbit telemetry data anomaly detection system, characterized in that, include: The data acquisition unit is used to acquire on-orbit telemetry time series data of the satellite to be tested; The time series data decomposition unit is used to introduce multi-scale wavelet transform and use the time-frequency characteristics of multi-scale wavelet transform to decompose the satellite on-orbit telemetry time series data into multiple approximate sequences and detail sequences. The data reconstruction unit is used to assign weights to each approximate sequence and detail sequence. By calculating the similarity score, the attention weight of each approximate sequence and detail sequence is calculated. The hidden states of each neuron in the coding layer and their corresponding attention weights are weighted and summed to calculate the content vector. The content vector is combined with the current hidden state of the neurons in the decoding layer, and so on, to obtain the reconstructed sequence of satellite on-orbit telemetry time series data. An anomaly detection unit is used to compare the difference between the reconstructed sequence and the real data sequence, and output abnormal data when the difference is greater than a set threshold. The method of decomposing satellite on-orbit telemetry time-series data into multiple approximate sequences and detail sequences using the time-frequency characteristics of multi-scale wavelet transform includes the following steps: A three-level Daubechies wavelet transform is used to decompose the satellite on-orbit telemetry time series data structure, which is decomposed into the first-level approximate sequence AS1 and the detail sequence DS1. The AS1 component is further decomposed into the second-level approximate sequence AS2 and the detail sequence DS2, and so on to complete the three-level Daubechies wavelet decomposition. The third-level approximate sequence AS3 and the four sub-sequences of the three-level detail sequences DS1, DS2 and DS3 are retained. It also includes calculating an attention score, which involves the following steps: By focusing on the attention of the query and the key, that is, given a query, the relevance between the query and the key is calculated; Determine the most suitable value based on the relevance between the query and the key; The attention score is obtained by calculating the similarity between the query at each position in the input sequence and the key at all positions. The calculation of attention weights for each approximate sequence and detail sequence includes the following steps: By normalizing the attention score, it is transformed into a probability distribution in which the sum of the weights of all elements is 1, and this distribution is used as the weight coefficient of the corresponding position Value. The result of the attention mechanism is obtained by weighting and summing the normalized weight coefficients with the Value. The encoder and decoder are integrated with LSTM recurrent units as the basic structure, and weights are assigned to each approximate sequence and detail sequence through the attention mechanism in the Transformer layer. Before decoding, the Transformer layer obtains the hidden states of all sequences in the encoding stage through the calculation results of the attention mechanism; it calculates a similarity score based on the hidden states of all sequences in the encoding stage and the hidden states of the decoding layer; and it calculates the weights of the attention mechanism's calculation results using the softmax function. a ij ; in, a ij The element to be encoded represents the correlation between other elements and the element with higher correlation; the element with higher correlation has a higher weight. a ij Larger.

Images