A satellite telemetry data anomaly detection method, device and electronic equipment

Abstract

The application discloses a satellite telemetry data anomaly detection method and device and electronic equipment, and the method comprises the following steps: acquiring a target telemetry data time sequence, performing multi-scale time slicing processing on the target telemetry data time sequence based on a preset slice scale set, and generating a plurality of target sequence segments; for each target sequence segment, after linear embedding mapping and superposition position coding are performed on the target sequence segment, a sub-sequence segment corresponding to the target sequence segment is generated; each sub-sequence segment is input into a preset multi-layer perception hybrid encoder for feature extraction, and different scale features are obtained; the different scale features are input into a preset multi-scale convolution fusion module for cross-scale feature fusion, and fusion features are generated; and the fusion features are processed to generate an anomaly detection result of satellite telemetry data. The application improves the accuracy of satellite telemetry data anomaly detection.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and specifically relates to a method, device and electronic equipment for detecting anomalies in satellite telemetry data. Background Technology

[0002] Satellite telemetry data is an important source of information reflecting the operational status of satellite platforms and payloads. By detecting anomalies in satellite telemetry data, problems such as equipment degradation, malfunctions, environmental disturbances, and potential faults can be identified in a timely manner, thereby providing support for satellite on-orbit health management, fault early warning, and operation and maintenance decisions.

[0003] Currently, existing anomaly detection methods for satellite telemetry data mainly fall into two categories: traditional methods and deep learning methods. Traditional methods primarily include those based on statistical analysis, clustering, and classification. These methods are simple to implement and have low computational overhead, but they typically rely on manual feature design or prior knowledge, making them less adaptable to high-dimensional, multivariate, strongly nonlinear, and non-stationary satellite telemetry sequences, and difficult to accurately characterize anomaly patterns under complex conditions. Existing deep learning methods, when modeling satellite telemetry sequences, often lack the ability to uniformly represent local perturbation features and global evolutionary patterns.

[0004] Therefore, improving the accuracy of anomaly detection in satellite telemetry data is an urgent technical problem to be solved. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a method for detecting anomalies in satellite telemetry data.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution: A method for detecting anomalies in satellite telemetry data, comprising: The target telemetry data time series is acquired, and multi-scale time slicing is performed on the target telemetry data time series based on a preset slice scale set to generate multiple target sequence segments; wherein, the preset slice scale set includes multiple different preset slice scales; For each target sequence segment, after performing linear embedding mapping and superposition position encoding on the target sequence segment, sub-sequence segments corresponding to the target sequence segment are generated; Each subsequence segment is input into a preset multilayer perceptron hybrid encoder for feature extraction to obtain features at different scales. The preset multilayer perceptron hybrid encoder consists of multiple stacked hybrid blocks. Each hybrid block includes a channel mixer, an inter-sequence mixer, and an intra-sequence mixer. The channel mixer is used to capture the dependencies between different variable channels, the inter-sequence mixer is used to capture the temporal dependencies and global trends between segments, and the intra-sequence mixer is used to capture the local features within each segment. Features at different scales are input into a preset multi-scale convolutional fusion module to perform cross-scale feature fusion and generate fused features. The fused features are processed to generate anomaly detection results from satellite telemetry data.

[0007] Optionally, each subsequence segment is input into a pre-defined multilayer perceptron hybrid encoder for feature extraction to obtain features at different scales, including: Each subsequence segment is input into a preset multilayer perceptron hybrid encoder. The channel dimension of the subsequence segment is transformed by the multilayer perceptron through the channel mixer to generate the first intermediate feature. The inter-sequence dimension of the first intermediate feature is transformed by a multilayer perceptron through an inter-sequence mixer to generate the second intermediate feature; By using an intra-sequence mixer, the intra-sequence dimension of the second intermediate feature is transformed by a multilayer perceptron to generate features at different scales.

[0008] Optionally, the process of multilayer perceptron transformation includes: The input features are projected to a high-dimensional space using a pre-set up-dimensional projection matrix to obtain high-dimensional features; A nonlinear activation function is applied to the high-dimensional features to perform a nonlinear transformation, resulting in the transformed features. A preset dimensionality reduction projection matrix is ​​used to map the transformed features back to the original input dimension, resulting in the output features after transformation by a multilayer perceptron.

[0009] Optionally, the preset multi-scale convolutional fusion module adopts a depthwise separable convolutional structure, inputting features of different scales into the preset multi-scale convolutional fusion module for cross-scale feature fusion to generate fused features, including: Features at different scales are input into a preset multi-scale convolution fusion module to perform depth convolution operations, and the first convolution result is obtained. The first convolution result is subjected to pointwise convolution to obtain the second convolution result, which is then used as the fused feature.

[0010] Optionally, the fused features are processed to generate anomaly detection results from the satellite telemetry data, including: Flatten the fused features to convert the high-dimensional features into a one-dimensional vector representation; The normalized prediction result is obtained by applying a fully connected linear mapping to the one-dimensional vector representation. The normalized prediction results are inversely standardized to obtain the final prediction results after restoring the original physical dimensions. After scoring anomalies based on the final prediction results, anomaly detection results of satellite telemetry data are generated.

[0011] Optionally, after anomaly scoring based on the final prediction results, anomaly detection results for satellite telemetry data are generated, including: The degree of deviation between the final prediction result and the actual telemetry value is calculated to obtain the initial anomaly score sequence; After smoothing the initial anomaly score sequence using an exponentially weighted moving average, the processed anomaly score sequence is obtained. A pre-defined peak value model is used to model the tail distribution of the processed abnormal scoring sequence, and an adaptive dynamic threshold is obtained by estimating using the generalized Pareto distribution. For each time step in the anomaly scoring sequence, the anomaly score is compared with an adaptive dynamic threshold to obtain the comparison result; If the comparison results indicate that the anomaly score exceeds the adaptive dynamic threshold, the anomaly detection result is an abnormal state; if the comparison results indicate that the anomaly score does not exceed the adaptive dynamic threshold, the anomaly detection result is a normal state.

[0012] Optionally, the time series of target telemetry data is acquired, including: Acquire multivariate telemetry data collected during satellite operation in orbit and construct an initial telemetry data time series; The initial telemetry data time series is standardized to generate the target telemetry data time series.

[0013] The present invention also provides a satellite telemetry data anomaly detection device, the device comprising: The slicing module is used to acquire the time series of target telemetry data, and to perform multi-scale time slicing processing on the target telemetry data time series based on a preset slice scale set to generate multiple target sequence fragments; wherein, the preset slice scale set includes multiple different preset slice scales; The generation module is used to generate sub-sequence segments corresponding to each target sequence segment after performing linear embedding mapping and superposition position encoding on the target sequence segment. The feature extraction module is used to input each sub-sequence segment into a preset multilayer perceptron hybrid encoder for feature extraction to obtain features at different scales. The preset multilayer perceptron hybrid encoder consists of multiple stacked mixing blocks. Each mixing block includes a channel mixer, an inter-sequence mixer, and an intra-sequence mixer. The channel mixer is used to capture the dependencies between different variable channels, the inter-sequence mixer is used to capture the temporal dependencies and global trends between segments, and the intra-sequence mixer is used to capture the local features within each segment. The feature fusion module is used to input features of different scales into a preset multi-scale convolutional fusion module to perform cross-scale feature fusion and generate fused features. The processing module is used to process the fused features and generate anomaly detection results from the satellite telemetry data.

[0014] Optionally, the above feature extraction module is specifically used for: Each subsequence segment is input into a pre-defined multilayer perceptron hybrid encoder for feature extraction, yielding features at different scales, including: Each subsequence segment is input into a preset multilayer perceptron hybrid encoder. The channel dimension of the subsequence segment is transformed by the multilayer perceptron through the channel mixer to generate the first intermediate feature. The inter-sequence dimension of the first intermediate feature is transformed by a multilayer perceptron through an inter-sequence mixer to generate the second intermediate feature; By using an intra-sequence mixer, the intra-sequence dimension of the second intermediate feature is transformed by a multilayer perceptron to generate features at different scales.

[0015] Optionally, the feature extraction module described above is also used for: The input features are projected to a high-dimensional space using a pre-set up-dimensional projection matrix to obtain high-dimensional features; A nonlinear activation function is applied to the high-dimensional features to perform a nonlinear transformation, resulting in the transformed features. A preset dimensionality reduction projection matrix is ​​used to map the transformed features back to the original input dimension, resulting in the output features after transformation by a multilayer perceptron.

[0016] Optionally, the preset multi-scale convolutional fusion module adopts a depthwise separable convolutional structure, and the aforementioned feature fusion module is specifically used for: Features at different scales are input into a preset multi-scale convolution fusion module to perform depth convolution operations, and the first convolution result is obtained. The first convolution result is subjected to pointwise convolution to obtain the second convolution result, which is then used as the fused feature.

[0017] Optionally, the above processing module is specifically used for: Flatten the fused features to convert the high-dimensional features into a one-dimensional vector representation; The normalized prediction result is obtained by applying a fully connected linear mapping to the one-dimensional vector representation. The normalized prediction results are inversely standardized to obtain the final prediction results after restoring the original physical dimensions. After scoring anomalies based on the final prediction results, anomaly detection results of satellite telemetry data are generated.

[0018] Optionally, the above processing module is also used for: The degree of deviation between the final prediction result and the actual telemetry value is calculated to obtain the initial anomaly score sequence; After smoothing the initial anomaly score sequence using an exponentially weighted moving average, the processed anomaly score sequence is obtained. A pre-defined peak value model is used to model the tail distribution of the processed abnormal scoring sequence, and an adaptive dynamic threshold is obtained by estimating using the generalized Pareto distribution. For each time step in the anomaly scoring sequence, the anomaly score is compared with an adaptive dynamic threshold to obtain the comparison result; If the comparison results indicate that the anomaly score exceeds the adaptive dynamic threshold, the anomaly detection result is an abnormal state; if the comparison results indicate that the anomaly score does not exceed the adaptive dynamic threshold, the anomaly detection result is a normal state.

[0019] Optionally, the above-mentioned slicing module is specifically used for: Acquire multivariate telemetry data collected during satellite operation in orbit and construct an initial telemetry data time series; The initial telemetry data time series is standardized to generate the target telemetry data time series.

[0020] The present invention also provides an electronic device, the electronic device including a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement the above-described method for detecting anomalies in satellite telemetry data.

[0021] The present invention also provides a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the above-described method for detecting anomalies in satellite telemetry data.

[0022] This invention provides a method, apparatus, and electronic device for anomaly detection in satellite telemetry data. The method includes: acquiring a time series of target telemetry data; performing multi-scale time slicing processing on the target telemetry data time series based on a preset slice scale set to generate multiple target sequence segments; for each target sequence segment, performing linear embedding mapping and overlay position encoding to generate a sub-sequence segment corresponding to the target sequence segment; inputting each sub-sequence segment into a preset multilayer perceptron hybrid encoder for feature extraction to obtain features at different scales; inputting the features at different scales into a preset multi-scale convolutional fusion module for cross-scale feature fusion to generate fused features; and processing the fused features to generate anomaly detection results for satellite telemetry data. This invention improves the accuracy of anomaly detection in satellite telemetry data.

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a satellite telemetry data anomaly detection method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a process for acquiring a time series of target telemetry data provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a process for generating features at different scales provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a process for generating fusion features according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a process for generating anomaly detection results provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of another process for generating anomaly detection results provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the experimental results after removing some modules according to an embodiment of the present invention; Figure 8 This is a schematic diagram of experimental results after removing part of the mixer according to an embodiment of the present invention; Figure 9 This is a framework diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0026] As the complexity of space missions continues to increase and the scale of satellite constellations expands, the frequency, dimensionality, and cumulative volume of telemetry data acquisition are growing rapidly. This has led to anomaly detection tasks exhibiting characteristics of large data scale, strong variable coupling, and complex dynamic changes. Furthermore, satellites operate in complex space environments for extended periods, and telemetry sequences are often simultaneously affected by multiple factors, including equipment operating conditions, changes in the orbital environment, external disturbances, and measurement noise, exhibiting significant non-stationarity, time-varying characteristics, and multi-timescale features. Strong correlations and coupling relationships also exist between different telemetry variables; anomalies are often not abrupt changes in a single variable, but rather shifts in multiple variables across different timescales.

[0027] Therefore, how to effectively extract key temporal features from complex, multivariable, and long-term telemetry sequences and accurately identify abnormal states has become an important research direction in the field of satellite health management.

[0028] Traditional anomaly detection techniques have limited ability to characterize complex anomaly patterns when processing high-dimensional, non-stationary satellite telemetry data, making it difficult to extract dynamic correlations between high-dimensional variables. While existing deep learning time-series models possess strong expressive power, their high computational complexity limits their application in large-scale telemetry data scenarios. However, simple network structures alone cannot guarantee detection performance because telemetry data intertwines low-frequency trends of long-term equipment degradation with high-frequency fluctuations from instantaneous environmental disturbances. Furthermore, existing methods generally employ a single, fixed sliding window strategy, lacking the ability to adaptively represent multi-scale features. This prevents the simultaneous capture of localized, minute transient fluctuations and global long-term evolutionary patterns, resulting in incomplete feature extraction and a high risk of missed detections of hidden anomalies or false positives of normal fluctuations.

[0029] To improve the accuracy of anomaly detection in satellite telemetry data, this invention provides a method for anomaly detection in satellite telemetry data.

[0030] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0031] like Figure 1 As shown, Figure 1 This is a flowchart illustrating a satellite telemetry data anomaly detection method provided in an embodiment of the present invention, including: Step 101: Obtain the target telemetry data time series, and perform multi-scale time slicing processing on the target telemetry data time series based on the preset slice scale set to generate multiple target sequence fragments.

[0032] The target telemetry data time series can be generated by processing the collected raw telemetry data. In some optional embodiments, such as... Figure 2 As shown, Figure 2 This is a schematic diagram of a process for acquiring a time series of target telemetry data provided in an embodiment of the present invention, including: Step 201: Acquire multivariate telemetry data collected during the satellite's on-orbit operation and construct an initial telemetry data time series.

[0033] Step 202: Standardize the initial telemetry data time series to generate the target telemetry data time series.

[0034] Multivariate telemetry data can include various parameters such as voltage, current, temperature, attitude, and power consumption, and is typically collected continuously in chronological order. Based on multivariate telemetry data, an initial telemetry data time series can be constructed. ,in, Indicates the batch data volume. Indicates the length of the input time window. This indicates the number of variable channels for telemetry data.

[0035] Because different telemetry variables differ in physical dimensions and numerical ranges, to reduce the impact of scale differences on the model training process, the initial telemetry data time series can be standardized to generate the target telemetry data time series. Specifically, based on the first... The first initial telemetry data time series The mean of each variable channel over the entire time dimension with standard deviation The following formula is used to standardize it: ; in, This is the result after standardization. For the first In the initial telemetry data time series, the first... The variable channel in the first... The specific initial telemetry values ​​at each time step.

[0036] To simultaneously capture local short-term fluctuations and long-term trend changes, multi-scale time slicing processing can be performed on the target telemetry data time series based on a preset slice scale set, generating multiple target sequence segments. The preset slice scale set includes multiple different preset slice scales; for example, the preset slice scale set can be denoted as... n1, n2, and n3 are three different preset slice scales, for length For target telemetry data time series, a sliding window method can be used to generate [data]. A target sequence segment, specifically denoted as... , S This is a preset fixed value; for example, it can be set to 2.

[0037] Step 102: For each target sequence segment, perform linear embedding mapping and superposition position encoding on the target sequence segment to generate the sub-sequence segment corresponding to the target sequence segment.

[0038] In order to unify the feature representation of target sequence segments at different scales, each target sequence segment can be... Mapping to a unified linear embedding D The implicit space representation, denoted as , is an embedding representation. Learnable positional encodings can be further superimposed on the embedding representation. E By injecting temporal position information, the corresponding subsequence fragments are obtained. H, recorded as .

[0039] Step 103: Input each subsequence segment into a preset multilayer perceptron hybrid encoder for feature extraction to obtain features at different scales.

[0040] The pre-defined multilayer perceptron hybrid encoder consists of multiple stacked hybrid blocks. Each hybrid block includes a channel mixer, an inter-sequence mixer, and an intra-sequence mixer. The channel mixer captures the dependencies between different variable channels, the inter-sequence mixer captures the temporal dependencies and global trends between segments, and the intra-sequence mixer captures the local features within each segment. Residual connections and layer normalization structures can be introduced between the various hybridization processes to improve the stability of network training.

[0041] In some alternative embodiments, such as Figure 3 As shown, Figure 3 This is a schematic diagram of a process for generating features at different scales according to an embodiment of the present invention, including: Step 301: Input each sub-sequence segment into a preset multilayer perceptron hybrid encoder, and perform multilayer perceptron transformation on the channel dimension of the sub-sequence segment through a channel mixer to generate the first intermediate feature.

[0042] Step 302: Perform a multilayer perceptron transformation on the inter-sequence dimension of the first intermediate feature using an inter-sequence mixer to generate the second intermediate feature.

[0043] Step 303: Perform multilayer perceptron transformation on the intra-sequence dimension of the second intermediate feature using an intra-sequence mixer to generate features at different scales.

[0044] The process of multilayer perceptron transformation includes: projecting the input features to a high-dimensional space using a pre-set up-dimensional projection matrix to obtain high-dimensional features; applying a non-linear activation function to the high-dimensional features to perform a non-linear transformation to obtain the transformed features; and mapping the transformed features back to the original input dimension using a pre-set down-dimensional projection matrix to obtain the output features after multilayer perceptron transformation. This bottleneck structure design of first expanding and then compressing enhances the model's ability to express complex dependencies through non-linear transformation in high-dimensional space.

[0045] Specifically, for any input tensor , M The length of the input tensor, Given the feature dimension of the input tensor, the computation process of the basic multilayer perceptron (MLP) is defined by the following formula: ; Among them, MLP ( XThe result is the calculation result of MLP; To predetermine the upgraded projection matrix, the expansion factor is set to... This makes the hidden layer dimension Expand to ; Preset dimension reduction projection matrix; and All are bias vectors; Dropout is a random deactivation operation used to suppress model overfitting and improve the generalization ability of telemetry data.

[0046] The multilayer perceptron architecture uses GELU as the activation function. The GELU activation function has a smoother nonlinear response near the zero point, provides a continuously differentiable activation pattern, and helps stabilize gradient propagation. Its mathematical definition is shown in the following equation: ; Among them, GELU ( x The output of the activation function is shown below. x The input variables for the activation function; This is the cumulative distribution function of the standard normal distribution.

[0047] Each subsequence segment is input into a preset multilayer perceptron hybrid encoder. When the channel dimensions of the subsequence segments are transformed using a channel mixer, this step captures the dependencies between different sensor channels due to the significant multivariate coupling characteristics of satellite telemetry data. The operation is applied to the variable channel dimensions. By sharing weights among channels, the interaction mechanism between multiple variables is learned, and cross-variable coupling features are extracted as the first intermediate feature. The calculation formula is shown in the following formula: ; in, This is the first intermediate feature; A multilayer perceptron network for extracting coupling features between variables; LN Layer normalization is used to standardize subsequence fragments. H .

[0048] The first intermediate feature is transformed using a multilayer perceptron through a sequence mixer to improve its inter-sequence dimension. Sequence mixing allows the model to aggregate contextual information between different time steps, thereby effectively capturing long-term evolutionary patterns in telemetry data and obtaining the second intermediate feature. The calculation formula is shown below: ; in, This is the second intermediate feature; For multilayer perceptron networks that act on the inter-sequence dimension of the first intermediate feature.

[0049] By applying a multilayer perceptron transformation to the intra-sequence dimension of the second intermediate feature through an intra-sequence mixer, and independently performing nonlinear transformations on the intra-segment representation, the model's ability to express local waveform details is enhanced. This outputs a deep feature representation containing multi-scale spatiotemporal coupling information, ultimately enabling the model to be based on multiple deep feature representations at different scales. That is, features at different scales can be obtained. F The calculation formula is shown below: ; in, This is the final deep feature representation; For multilayer perceptron networks that act on the intra-sequence dimension of the second intermediate feature.

[0050] Step 104: Input features of different scales into the preset multi-scale convolution fusion module to perform cross-scale feature fusion and generate fused features.

[0051] After feature extraction, features at different scales can be input into a preset multi-scale convolutional fusion module for cross-scale feature fusion, thereby generating fused features. In some optional embodiments, such as... Figure 4 As shown, Figure 4 This is a schematic diagram of a process for generating fusion features according to an embodiment of the present invention, including: Step 401: Input features of different scales into the preset multi-scale convolution fusion module to perform depth convolution operation and obtain the first convolution result.

[0052] Step 402: Perform pointwise convolution on the first convolution result to obtain the second convolution result, and use the second convolution result as the fusion feature.

[0053] Among them, the preset multi-scale convolution fusion module adopts a depthwise separable convolution structure. The depthwise separable convolution structure models the interaction relationship between features at different scales through hierarchical convolution operations, so that feature representations at different time scales can form a collaborative expression.

[0054] Specifically, the following formula can be used to perform depthwise convolution operations on features of different scales to obtain the first convolution result. F 1: ; in, This is a depthwise convolution operation.

[0055] Next, the first convolution result can be subjected to pointwise convolution operation using the following formula to obtain the second convolution result, and the second convolution result can be used as the fused feature. F 2: ; in, This is a pointwise convolution operation.

[0056] In this embodiment, the two convolutions are two steps of depthwise separable convolution, the purpose of which is to compute the first... When outputting features at each scale, the encoded features of the current scale are fused with the output features of adjacent scales, thereby establishing a direct information transfer relationship between different scales. Through this connection method, each scale feature, while maintaining its own structural information, can receive supplementary information from other scales step by step, enabling feature representations at different time scales to form a collaborative expression.

[0057] Step 105: Process the fused features to generate anomaly detection results for satellite telemetry data.

[0058] After generating the fusion features, the fusion features can be processed to generate corresponding anomaly detection results. In some optional embodiments, such as... Figure 5 As shown, Figure 5 This is a schematic diagram of a process for generating anomaly detection results provided by an embodiment of the present invention, including: Step 501: Flatten the fused features to convert the high-dimensional features into a one-dimensional vector representation.

[0059] Step 502: Obtain the normalized prediction result by mapping the one-dimensional vector representation through a fully connected linear mapping.

[0060] Step 503: Perform inverse standardization on the normalized prediction results to obtain the final prediction results after restoring the original physical dimensions.

[0061] Step 504: After scoring the anomalies based on the final prediction results, generate the anomaly detection results of the satellite telemetry data.

[0062] After completing multi-scale feature fusion, the fused features can be... F 2. Perform a flattening operation to convert the high-dimensional features into a one-dimensional vector representation. Y Then represent it as a one-dimensional vector. Y Normalized prediction results are obtained through fully connected linear mapping. Then, the normalized prediction results are inversely standardized using the statistics recorded during the standardization phase to restore the original physical dimensions, thus obtaining the final prediction result after restoring the original physical dimensions. This can be calculated using the following formula: ; ; ; in, Indicates the flattening operation; This is the prediction layer weight matrix; It is the bias vector; This is the final prediction result; and These are the mean and standard deviation, which are the values ​​saved at the end of the standardization process.

[0063] In some alternative embodiments, such as Figure 6 As shown, Figure 6 This is a schematic diagram of another process for generating anomaly detection results provided by an embodiment of the present invention, including: Step 601: Calculate the degree of deviation between the final prediction result and the actual telemetry value to obtain the initial anomaly score sequence.

[0064] Step 602: After smoothing the initial abnormal score sequence by exponential weighted moving average, the processed abnormal score sequence is obtained.

[0065] Step 603: The tail distribution of the processed abnormal scoring sequence is modeled using a preset over-threshold peak model, and the adaptive dynamic threshold is estimated using the generalized Pareto distribution.

[0066] Step 604: For the abnormal score at each time step in the abnormal score sequence, compare the abnormal score with the adaptive dynamic threshold to obtain the comparison result.

[0067] Step 605: If the comparison result determines that the abnormal score exceeds the adaptive dynamic threshold, the abnormal detection result is an abnormal state; if the comparison result determines that the abnormal score does not exceed the adaptive dynamic threshold, the abnormal detection result is a normal state.

[0068] The following formula can be used to calculate the degree of deviation between the final prediction result and the actual telemetry value, and obtain the initial anomaly score sequence: ; in, This is the initial anomaly scoring sequence; These are the actual telemetry values; These are the weighting coefficients corresponding to each variable channel.

[0069] Next, a pre-defined Peaks Over Threshold (POT) model is used to model the tail distribution of the processed anomaly score sequence, and an adaptive dynamic threshold is estimated using a generalized Pareto distribution. Specifically, assuming the tail distribution of the anomaly score sequence follows a generalized Pareto distribution, after obtaining its distribution function, the shape and scale parameters of the generalized Pareto distribution can be fitted using maximum likelihood estimation. Then, an anomaly detection significance level can be set, and the corresponding quantiles are calculated based on the fitted shape and scale parameters. This quantile is then used as the portion exceeding the pre-defined initial threshold. Adding this quantile to the pre-defined initial threshold yields the threshold, which is updated with new anomaly scores to achieve adaptive dynamic adjustment, thus obtaining the aforementioned adaptive dynamic threshold.

[0070] In this embodiment, by combining anomaly scoring based on prediction residuals with the POT dynamic threshold determination mechanism, the model can achieve adaptive anomaly detection without the need for anomaly labels.

[0071] Furthermore, this application embodiment further illustrates the point through the following simulation experiments. The experiments specifically utilize publicly released standard open-source satellite anomaly datasets, including the Soil Moisture Active Passive (SMAP) satellite telemetry dataset and the Mars Science Laboratory (MSL) satellite telemetry dataset. Both datasets originate from real-world space mission scenarios and can effectively reflect the diverse anomaly types and complex evolutionary forms in satellite telemetry data. The anomaly forms in the SMAP and MSL satellite telemetry datasets encompass point anomalies, ensemble anomalies, and contextual anomalies, placing high demands on the anomaly detection algorithms in terms of sensitivity and stability.

[0072] The SMAP satellite telemetry dataset consists of 8 anomalous event subsequences, each containing 25-dimensional telemetry features, including 13 point anomalies, 40 ensemble anomalies, and 19 context anomalies. The MSL satellite telemetry dataset contains 19 anomalous event subsequences, each represented by 55-dimensional telemetry features, including 3 point anomalies, 30 ensemble anomalies, and 3 context anomalies. Experiments were conducted for comparison and validation using the same training and test sets. In model training, the input sequences were first preprocessed by standardization and multi-scale segmentation, using mean squared error as the loss function between predicted and true values. During training, the Adam optimizer was used for backpropagation to iteratively update network parameters until the network model loss no longer decreased and tended to stabilize, thus obtaining the optimal anomaly detection model.

[0073] To evaluate the effectiveness of the proposed lightweight satellite telemetry data anomaly detection algorithm based on multi-scale feature fusion, this invention selects Local Outlier Factor (LOF), Isolation Forest (IForest), Long Short-Term Memory Variational Autoencoder (LSTM-VAE), Omni-Dimensional Anomaly Detection (OmniAnomaly), Anomaly Transformer (AnomalyTrans), and TimesNet as comparison algorithms. MSPatch-Mixer is the multi-scale segment mixer method in this invention, which employs multi-scale time slicing processing and a pre-set multilayer perceptron hybrid encoder. Experiments were conducted on the SMAP and MSL satellite telemetry datasets. The detection performance is shown in Table 1. Evaluation metrics include precision, recall, and F1 score.

[0074] Table 1

[0075] The results show that on the SMAP dataset, the F1 score of the method of this invention reaches 93.56%, which is about 32% and 27% higher than LOF and IFOreest, respectively, and also outperforms methods such as LSTM-VAE, OmniAnomaly, and TimesNet. On the MSL dataset, the F1 score of the method of this invention reaches 91.26%, which is about 34% and 36% higher than LOF and IFOreest, respectively, and about 22% higher than TimesNet. The overall detection performance is better than most of the comparison models.

[0076] To verify the impact of each component module and multi-scale structure of the method of this invention on anomaly detection performance, ablation experiments and scale combination experiments were conducted, and the results are as follows: Figure 7 and Figure 8 As shown. Figure 7 This is a schematic diagram of the experimental results after removing some modules, provided by an embodiment of the present invention. Figure 8 This is a schematic diagram of the experimental results after removing part of the mixer, provided by an embodiment of the present invention.

[0077] Figure 7The Patch MLP-Mixer structure is a pre-defined multilayer perceptron hybrid encoder, and the HPC fusion module is a pre-defined multi-scale convolutional fusion module. Experimental results show that when multi-scale features, the Patch MLP-Mixer structure, or the HPC fusion module are removed, the model's precision, recall, and F1 score all decrease to varying degrees. The performance degradation is most significant when the Patch MLP-Mixer structure is removed, indicating that this structure plays a key role in temporal feature modeling.

[0078] Figure 8 The effects of the three mixing mechanisms were further analyzed: Channel-Mixer (channel mixer), Inter-Mixer (inter-sequence mixer), and Intra-Mixer (intra-sequence mixer). The results showed that removing any module from Channel-Mixer, Inter-Mixer, or Intra-Mixer would lead to a decrease in model performance, indicating that the three mixing mechanisms have complementary effects in variable association modeling, cross-temporal dependency modeling, and local feature extraction.

[0079] This invention proposes a target tracking method based on token convolution. The method includes: disclosing a satellite telemetry data anomaly detection method, device, and electronic device; the method includes: acquiring a target telemetry data time series; performing multi-scale time slicing processing on the target telemetry data time series based on a preset slice scale set to generate multiple target sequence segments; for each target sequence segment, performing linear embedding mapping and superimposed position encoding to generate a corresponding sub-sequence segment; inputting each sub-sequence segment into a preset multi-layer perceptron hybrid encoder for feature extraction to obtain features at different scales; inputting the features at different scales into a preset multi-scale convolution fusion module for cross-scale feature fusion to generate fused features; and processing the fused features to generate anomaly detection results for satellite telemetry data. This invention combines multi-scale time series modeling with a lightweight deep network structure, integrating multi-scale feature extraction, temporal feature encoding, and anomaly detection processes through an end-to-end framework to achieve efficient identification of complex telemetry anomaly patterns. Furthermore, by introducing a pre-defined multilayer perceptron hybrid encoder and a pre-defined multi-scale convolutional fusion module, it jointly models variable coupling relationships and temporal dependencies through channel mixing, inter-sequence mixing, and intra-sequence mixing mechanisms. This reduces model complexity while maintaining detection performance, improving computational efficiency and real-time detection capabilities. Simultaneously, it utilizes a multi-scale slicing strategy to extract feature information at different time scales, thereby enhancing the model's ability to characterize local instantaneous fluctuations and long-term evolution trends, improving the accuracy of anomaly detection in satellite telemetry data. In addition, this invention significantly reduces model computational complexity and resource consumption while maintaining anomaly detection accuracy, thus improving anomaly detection performance and real-time application capabilities in large-scale satellite telemetry data scenarios.

[0080] Based on the same inventive concept, embodiments of the present invention also provide a satellite telemetry data anomaly detection device, the device comprising: The slicing module is used to acquire the time series of target telemetry data, and to perform multi-scale time slicing processing on the target telemetry data time series based on a preset slice scale set to generate multiple target sequence fragments; wherein, the preset slice scale set includes multiple different preset slice scales.

[0081] The generation module is used to generate sub-sequence segments corresponding to each target sequence segment after performing linear embedding mapping and superposition position encoding on the target sequence segment.

[0082] The feature extraction module is used to input each sub-sequence segment into a preset multilayer perceptron hybrid encoder for feature extraction to obtain features at different scales. The preset multilayer perceptron hybrid encoder consists of multiple stacked mixing blocks. Each mixing block includes a channel mixer, an inter-sequence mixer, and an intra-sequence mixer. The channel mixer is used to capture the dependencies between different variable channels, the inter-sequence mixer is used to capture the temporal dependencies and global trends between segments, and the intra-sequence mixer is used to capture the local features within each segment.

[0083] The feature fusion module is used to input features of different scales into a preset multi-scale convolution fusion module to perform cross-scale feature fusion and generate fused features.

[0084] The processing module is used to process the fused features and generate anomaly detection results from the satellite telemetry data.

[0085] This invention also provides an electronic device, such as... Figure 9 As shown, it includes a processor 901, a communication interface 902, a memory 903, and a communication bus 904. The processor 901, communication interface 902, and memory 903 communicate with each other via the communication bus 904. Memory 903 is used to store computer programs; The processor 901, when executing the program stored in the memory 903, implements the steps of the above-mentioned method for detecting anomalies in satellite telemetry data.

[0086] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0087] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0088] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0089] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0090] The present invention also provides a computer-readable storage medium. A computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program implements the steps of the above-described method for detecting anomalies in satellite telemetry data.

[0091] Optionally, the computer-readable storage medium may be non-volatile memory (NVM), such as at least one disk storage device.

[0092] Optionally, the computer-readable storage medium may also be at least one storage device located remotely from the aforementioned processor.

[0093] It should be noted that, for the embodiments of the device / electronic device / storage medium, since they are basically similar to the method embodiments, the description is relatively simple. For relevant parts, please refer to the description of the method embodiments. All embodiments of the above-mentioned satellite telemetry data anomaly detection method are applicable to the device, electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0094] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for detecting anomalies in satellite telemetry data, characterized in that, The method includes: A time series of target telemetry data is acquired, and multi-scale time slicing is performed on the target telemetry data time series based on a preset slice scale set to generate multiple target sequence segments; wherein, the preset slice scale set includes multiple different preset slice scales; For each target sequence segment, after performing linear embedding mapping and superposition position encoding on the target sequence segment, a sub-sequence segment corresponding to the target sequence segment is generated; Each of the subsequence segments is input into a preset multilayer perceptron hybrid encoder for feature extraction to obtain features at different scales. The preset multilayer perceptron hybrid encoder consists of multiple stacked hybrid blocks. Each hybrid block includes a channel mixer, an inter-sequence mixer, and an intra-sequence mixer. The channel mixer is used to capture the dependencies between different variable channels. The inter-sequence mixer is used to capture the temporal dependencies and global trends between segments. The intra-sequence mixer is used to capture the local features within each segment. The features at different scales are input into a preset multi-scale convolutional fusion module to perform cross-scale feature fusion and generate fused features. The fused features are processed to generate anomaly detection results for satellite telemetry data.

2. The method according to claim 1, characterized in that, The process involves inputting each of the sub-sequence segments into a preset multilayer perceptron hybrid encoder for feature extraction to obtain features at different scales, including: Each of the subsequence segments is input into a preset multilayer perceptron hybrid encoder, and the channel dimension of the subsequence segments is transformed by the channel mixer to generate a first intermediate feature; The inter-sequence dimension of the first intermediate feature is transformed by a multilayer perceptron through the inter-sequence mixer to generate the second intermediate feature; The intra-sequence dimension of the second intermediate feature is transformed by a multilayer perceptron through the intra-sequence mixer to generate the features at different scales.

3. The method according to claim 2, characterized in that, The process of the multilayer perceptron transformation includes: The input features are projected to a high-dimensional space using a pre-set up-dimensional projection matrix to obtain high-dimensional features; A nonlinear activation function is applied to the high-dimensional features to perform a nonlinear transformation, resulting in the transformed features. The transformed features are mapped back to the original input dimension using a preset dimension reduction projection matrix to obtain the output features after transformation by a multilayer perceptron.

4. The method according to any one of claims 1-3, characterized in that, The preset multi-scale convolutional fusion module adopts a depthwise separable convolutional structure. The step of inputting the features of different scales into the preset multi-scale convolutional fusion module for cross-scale feature fusion to generate fused features includes: The features at different scales are input into a preset multi-scale convolution fusion module to perform a depth convolution operation, and the first convolution result is obtained. Perform a pointwise convolution operation on the first convolution result to obtain a second convolution result, and use the second convolution result as the fusion feature.

5. The method according to any one of claims 1-3, characterized in that, The process of processing the fused features to generate anomaly detection results for satellite telemetry data includes: The fused features are flattened to convert the high-dimensional features into a one-dimensional vector representation; The one-dimensional vector representation is then mapped through a fully connected linear method to obtain a normalized prediction result. The normalized prediction results are inversely normalized to obtain the final prediction results after restoring the original physical dimensions. After performing anomaly scoring based on the final prediction results, anomaly detection results of the satellite telemetry data are generated.

6. The method according to claim 5, characterized in that, After performing anomaly scoring based on the final prediction result, the generation of anomaly detection results for the satellite telemetry data includes: The degree of deviation between the final prediction result and the actual telemetry value is calculated to obtain an initial anomaly score sequence; After applying an exponentially weighted moving average smoothing to the initial anomaly score sequence, the processed anomaly score sequence is obtained. A preset over-threshold peak model is used to model the tail distribution of the processed abnormal scoring sequence, and an adaptive dynamic threshold is estimated using a generalized Pareto distribution. For each time step in the abnormal scoring sequence, the abnormal score is compared with the adaptive dynamic threshold to obtain the comparison result; If the anomaly score exceeds the adaptive dynamic threshold based on the comparison result, the anomaly detection result is an abnormal state; if the anomaly score does not exceed the adaptive dynamic threshold based on the comparison result, the anomaly detection result is a normal state.

7. The method according to any one of claims 1-3, characterized in that, The acquisition of the target telemetry data time series includes: Acquire multivariate telemetry data collected during satellite operation in orbit and construct an initial telemetry data time series; The initial telemetry data time series is standardized to generate the target telemetry data time series.

8. A satellite telemetry data anomaly detection device, characterized in that, The device includes: The slicing module is used to acquire the time series of target telemetry data, and to perform multi-scale time slicing processing on the time series of target telemetry data based on a preset slice scale set to generate multiple target sequence fragments; wherein, the preset slice scale set includes multiple different preset slice scales; The generation module is used to generate sub-sequence segments corresponding to each target sequence segment by performing linear embedding mapping and superposition position encoding on the target sequence segments. The feature extraction module is used to input each of the sub-sequence segments into a preset multilayer perceptron hybrid encoder for feature extraction to obtain features at different scales. The preset multilayer perceptron hybrid encoder is composed of multiple stacked mixing blocks. Each mixing block includes a channel mixer, an inter-sequence mixer, and an intra-sequence mixer. The channel mixer is used to capture the dependencies between different variable channels. The inter-sequence mixer is used to capture the temporal dependencies and global trends between segments. The intra-sequence mixer is used to capture the local features within each segment. The feature fusion module is used to input the features at different scales into a preset multi-scale convolutional fusion module to perform cross-scale feature fusion and generate fused features. The processing module is used to process the fused features and generate anomaly detection results for the satellite telemetry data.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the satellite telemetry data anomaly detection method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the satellite telemetry data anomaly detection method as described in any one of claims 1-7.

Images