Satellite-borne product anomaly detection method based on multi-dimensional telemetering time sequence STL decomposition

Abstract

The invention relates to a satellite-borne product anomaly detection method based on multi-dimensional telemetering time sequence STL decomposition. The method comprises the following steps: constructing a standardized time sequence data table; asynchronous multi-channel alignment and down-sampling are completed through a fixed time window, data with different sampling frequencies and time offsets are mapped to a unified time axis, and a multi-channel time sequence matrix is generated. Then STL time sequence decomposition is carried out on the matrix, three independent components of a trend, a period and a residual error are decoupled, and differential modeling is carried out by respectively adopting an LSTM auto-encoder, a 1D-CNN auto-encoder and statistical distribution; and finally, calculating each model reconstruction error and performing weighted fusion to obtain a comprehensive anomaly score, constructing a threshold value based on normal sample distribution to complete point-by-point anomaly judgment, and positioning an anomaly source according to each component error weight distribution, thereby efficiently solving the problems of multi-channel asynchronization, strong noise and the like, and realizing high-precision anomaly detection.

Description

Technical Field

[0001] This application relates to the fields of spacecraft health management and satellite on-orbit intelligent monitoring technology, and in particular to a method for detecting anomalies in spaceborne products based on multidimensional telemetry time series STL decomposition. Background Technology

[0002] As a core component of spacecraft, the on-orbit operational status of spaceborne products directly determines the success or failure of space missions. Anomaly detection for spaceborne products primarily relies on telemetry data, analyzing multi-channel time-series parameters such as voltage, current, and temperature to identify precursory faults and abnormal states. However, existing technologies face numerous bottlenecks in practical applications, making it difficult to meet the high-precision detection requirements of complex spaceborne environments. First, spaceborne telemetry data suffers from multi-channel asynchronous sampling and strong noise interference. The sampling frequencies of different sensors vary significantly, and time shifts exist, making it difficult to directly couple and analyze multi-dimensional data. Simultaneously, factors such as space electromagnetic radiation and inherent sensor errors introduce a large amount of high-frequency random noise, severely impacting data quality. Traditional methods often employ simple interpolation alignment or direct data stitching, which fails to solve the timing synchronization problem and effectively suppresses noise, resulting in a weak foundation for modeling.

[0003] Secondly, spaceborne telemetry signals contain mixed time-scale features such as long-term slow-changing trends, short-term periodic fluctuations, and random residuals. Traditional end-to-end deep learning models (such as single LSTM and CNN) adopt a "one-size-fits-all" modeling approach, making it difficult to capture features at different scales simultaneously. Subtle long-term aging trends are often masked by large-scale periodic fluctuations or noise, resulting in insufficient sensitivity of the model to latent degradation features and an inability to achieve early fault warning.

[0004] Furthermore, abnormal samples of onboard products are extremely scarce, and most of their operation is in a normal state. Traditional deep neural networks tend to forcibly fit Gaussian white noise during training, leading to an increased risk of overfitting and a high false alarm rate. Existing methods mostly mitigate this through data augmentation or regularization, but they do not differentiate on-orbit onboard products from the modeling mechanism. As crucial units providing functional support within the satellite system, their long-term stability directly impacts the overall performance of the satellite. In complex space environments, onboard products are affected by various factors such as changes in radiation environment, fluctuations in thermal control conditions, and device aging, which may lead to gradual performance degradation, such as output parameter drift and anomalies. Traditional on-orbit health monitoring methods are usually based on fixed thresholds for single telemetry indicators, making it difficult to identify hidden performance degradation trends or capture correlational anomalies among multiple parameters.

[0005] Currently, the main technical challenges in health monitoring of on-orbit operation of spaceborne products are as follows: (1) Multi-channel asynchronous sampling. Different measurement links have different sampling periods and sampling offsets, and the timestamps of multiple channels differ. Due to the different clock synchronization mechanisms between the control system and the payload, it is difficult to directly and uniformly model multi-channel data.

[0006] (2) Abnormal samples are extremely scarce. There are very few cases of on-orbit failures of spaceborne products, and there is almost no sufficient labeled abnormal data for supervised learning. Traditional classification anomaly detection methods based on supervised learning are difficult to apply.

[0007] (3) Long-term gradual trends are difficult to capture. The aging and performance changes of spaceborne products usually manifest as a slow drift that lasts for months or even years. Short time windows cannot effectively perceive this long-term trend, while using long windows will cause problems such as excessive computational overhead, long training time, and insufficient real-time performance.

[0008] (4) Complex coupling relationships among multidimensional parameters. Spaceborne products contain multiple key telemetry parameters, and there are complex physical coupling relationships among these parameters. Single-parameter threshold monitoring cannot detect latent anomalies of multi-parameter collaborative degradation.

[0009] (5) Severe noise interference. Telemetry time series data contains large-amplitude white noise and various random interferences. If directly input into a deep learning model for modeling, the model is likely to learn noise features rather than the real physical change patterns, thereby reducing the effectiveness of anomaly detection indicators.

[0010] In existing technologies, detection methods based on fixed thresholds heavily rely on human experience, can only identify obvious sudden anomalies, cannot capture latent degradation trends, and are sensitive to noise with high false alarm and false negative rates. End-to-end detection methods based on deep learning do not distinguish between features of different time scales, are susceptible to data gaps and non-uniform sampling, have complex models leading to a high risk of overfitting, poor interpretability, and are difficult to adapt to the complex characteristics of spaceborne telemetry data. Therefore, there is an urgent need for a new technical solution that can adapt to the characteristics of spaceborne product telemetry data and achieve high-precision anomaly detection under conditions of scarce anomaly samples. Summary of the Invention

[0011] Therefore, it is necessary to provide a spaceborne product anomaly detection method based on multidimensional telemetry time series STL decomposition that can adapt to the segmented, irregular, missing, multidimensional coupling, and long-term slow-varying characteristics of telemetry data of spaceborne products and improve the accuracy of anomaly detection, in order to address the above-mentioned technical problems.

[0012] A method for detecting anomalies in spaceborne products based on multidimensional telemetry time-series STL decomposition, the method comprising: A standardized time-series data table is constructed by loading and structuring telemetry data of multi-channel spaceborne products; asynchronous multi-channel alignment and downsampling are performed on the standardized time-series data based on a fixed time window, and multi-channel data with different sampling frequencies and time offsets are mapped to a unified time axis to obtain a window-aligned multi-channel time-series matrix. STL time series decomposition is performed on the multi-channel time series matrix to decouple the complex telemetry signal into three independent components: trend, period, and residual. Differential independent modeling is performed on the three independent components: LSTM autoencoder is used for the trend component, 1D-CNN autoencoder is used for the period component, and statistical distribution is used for the residual component. The reconstruction errors of each model obtained by computational modeling are weighted and fused to obtain a comprehensive anomaly score by combining the reconstruction errors of the trend component, the reconstruction errors of the periodic component, and the deviation of the residual component. A judgment threshold is constructed based on the normal sample distribution. The anomaly score and the threshold are combined to complete the point-by-point anomaly judgment. At the same time, the source of the anomaly is located according to the error weight distribution of each component.

[0013] The aforementioned method for anomaly detection in spaceborne products based on multidimensional telemetry time-series STL decomposition first constructs a standardized time-series data table through multi-channel telemetry data loading and structured processing, laying a standardized data foundation for subsequent processing. The asynchronous multi-channel alignment and downsampling mechanism based on a fixed time window not only maps multidimensional data with different sampling frequencies and time offsets to a unified time axis, solving the time-series synchronization problem in multidimensional coupled analysis, but also utilizes the window mean to naturally suppress high-frequency random noise. Simultaneously, it handles irregular missing data through interpolation or pre- and post-filling, adapting to the characteristics of data segmentation. Secondly, the innovative STL time-series decomposition strategy decouples complex telemetry signals into three independent components: trend, period, and residual. This fundamentally avoids long-term, slowly changing trends being masked by short-term periodic fluctuations or noise, allowing subtle aging and degradation characteristics to be highlighted, significantly improving sensitivity to latent anomalies. Furthermore… This approach employs differentiated and independent modeling for the three components, taking into account their distinct characteristics. An LSTM autoencoder accurately captures the long-range dependence of the trend component, a 1D-CNN autoencoder effectively extracts the local structure of the periodic component, and statistical distribution modeling adapts to the noise characteristics of the residual component. This design avoids the forced fitting of deep models to Gaussian white noise when outlier samples are extremely scarce, thus mitigating the risk of overfitting and reducing the false alarm rate. Finally, a comprehensive outlier score is obtained by weighted fusion of the errors of the three components. An adaptive threshold is constructed by combining the distribution of normal samples, enabling accurate capture of all types of anomalies, including instantaneous outliers, progressive degradation, and waveform distortion. Furthermore, the error weight distribution of each component allows for intuitive location of the anomaly source, providing strong interpretability of the detection results and further ensuring the accuracy and reliability of anomaly detection. This approach comprehensively adapts to the complex characteristics of telemetry data from spaceborne products. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a method for detecting anomalies in spaceborne products based on multidimensional telemetry time-series STL decomposition in one embodiment. Figure 2 This is a schematic diagram of multi-dimensional time-series downsampling in one embodiment; Figure 3 This is a schematic diagram of timing STL decomposition in one embodiment; Figure 4 Here is a schematic diagram of the LSTM autoencoder results and LSTM unit in another embodiment; Figure 5 This is a schematic diagram of the original telemetry time series, downsampling time series, decomposed trend term time series, period term time series, residual term time series, anomaly score time series, and detection result time series of normal data and outlier data in one embodiment. Figure 6 This is a schematic diagram of the original telemetry time series, downsampling time series, decomposed trend term time series, period term time series, residual term time series, anomaly score time series, and detection result time series of the overall drift data and slope change data in one embodiment; Figure 7 This is a schematic diagram of the original telemetry time series, downsampling time series, decomposed trend term time series, period term time series, residual term time series, anomaly score time series, and detection result time series of mixed abnormal data in one embodiment. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0016] In one embodiment, such as Figure 1 As shown, a method for anomaly detection of spaceborne products based on multidimensional telemetry time series STL decomposition is provided, including the following steps: Step 102: Load and structure the telemetry data of the multi-channel spaceborne products to construct a standardized time series data table; perform asynchronous multi-channel alignment and downsampling on the standardized time series data based on a fixed time window, and map the multi-channel data with different sampling frequencies and different time offsets to a unified time axis to obtain the multi-channel time series matrix after window alignment.

[0017] Multiple telemetry channels related to the onboard products are acquired to form raw time-series data. The raw downlink telemetry frames undergo quality assessment and filtering. First, data rows with parsed error states are removed, retaining only regular downlink frames with consistent sampling intervals, while frames with abnormal sampling intervals or non-service sampling are removed. The dispersed multi-channel telemetry data is organized into a standardized format, and invalid data is removed, providing a high-quality data foundation for subsequent time-series alignment, decomposition, and modeling. Specifically, multi-channel onboard product telemetry data refers to operational status parameter data collected by different sensors from key satellite components, including various physical quantities characterizing the product's operating status such as voltage, current, temperature, frequency, and phase. The standardized time-series data table is a structured data carrier containing timestamps, satellite identifiers, and telemetry values ​​from each channel, facilitating subsequent unified processing.

[0018] To address the asynchronous sampling characteristics of multi-channel telemetry data, the earliest timestamp of all channels is used first. Establish a unified starting point for windows:

[0019] With fixed window length (e.g., 30 minutes) Generate continuous windows:

[0020] This window sequence serves as a unified timeline for all channels, enabling multi-channel data to be aligned on the same set of windows.

[0021] like Figure 2 As shown, the process of downsampling a multidimensional time series window includes: For each channel , sampling time Mapped to the corresponding window And construct a set of data points within the window:

[0022] Calculate the window mean based on this:

[0023] If there is no data in the window, the value is generated by interpolation or padding before and after.

[0024] For windows The window mean values ​​of all channels are combined into a vector:

[0025] This yields the window-aligned multi-channel timing matrix:

[0026] This time series matrix eliminates timestamp differences and serves as a unified input format for STL decomposition and subsequent deep learning models.

[0027] A unified window starting point is determined based on the earliest timestamp of all channels. A continuous time window sequence is generated with a preset fixed window length. This window sequence serves as a unified time axis for all channels, providing a benchmark for multi-channel data alignment. The sampling time of each channel is mapped to the corresponding time window. The effective data points within each window are counted, and the window mean is calculated as the representative value of that window for that channel. If there is no data within a window, it is supplemented by interpolation or padding before and after, achieving data downsampling and noise suppression. The window means of all channels under each time window are combined into a vector to form a multi-channel time series matrix after window alignment. This matrix eliminates the timestamp differences of multi-channel data, providing a unified input for subsequent STL decomposition and deep learning modeling.

[0028] Step 104: Perform STL time series decomposition on the multi-channel time series matrix to decouple the complex telemetry signal into three independent components: trend, period, and residual. Perform differentiated independent modeling for the three independent components: LSTM autoencoder for trend component, 1D-CNN autoencoder for period component, and statistical distribution modeling for residual component.

[0029] Based on the system's known master period (e.g., 13 hours) and the time resolution after window alignment (e.g., 30 minutes), the number of windows contained in each period can be obtained:

[0030] For example, when the period Window length ,but:

[0031] This parameter serves as the seasonal window input for STL, enabling periodic decomposition to accurately pinpoint the inherent oscillation structure of the onboard product. Before performing STL, missing windows are linearly interpolated, and edge regions are filled in to ensure signal continuity and decomposability.

[0032] Figure 3 The diagram shown illustrates the temporal STL decomposition, where the STL decomposition divides the window sequence... Decomposed into:

[0033] in, The slow change trend reflects device aging and long-term drift. As a periodic term, it reflects controlled, regular behavior or operational rhythm. It is a high-frequency residual, which includes sudden disturbances, noise, and local anomalies.

[0034] STL uses a three-stage iterative algorithm based on LOESS local regression, which includes: smoothing the periodic structure by using LOESS on a periodic window to ensure that the periodic signal has a stable shape in different windows; applying a large window LOESS to extract the gradual trend based on the deperiodic signal, so that the model can capture the decay or aging trend on a multi-day to multi-week scale; and obtaining residuals through differencing to localize random disturbances and anomalies, which serve as the most sensitive subspace for anomaly detection.

[0035] To address high noise and outliers, this step enables robust mode in STL, using iterative weights to suppress the pseudo-influence of outliers on trends and cycles, making the decomposition more stable and reliable.

[0036] For each channel Perform STL decomposition independently to obtain

[0037]

[0038]

[0039] Each column represents a sequence corresponding to a specific sensor, and each row corresponds to a time window. .

[0040] The number of windows in each period is calculated based on the known main period and window length of the system, and used as the seasonal window input for STL decomposition; linear interpolation is performed on the missing windows in the time series matrix to fill in the edge regions and ensure signal continuity and decomposability. A robust STL decomposition algorithm is adopted, which extracts components through a three-stage iteration: the first stage extracts stable periodic components by smoothing the LOESS over a periodic window; the second stage applies a large-window LOESS to the deperiodic signal to extract slowly varying trend components on a multi-day to multi-week scale; the third stage obtains residual components containing sudden disturbances, noise, and local anomalies through differential operations. During the iteration process, the weights are dynamically adjusted to suppress the interference of outliers on the fitting results, ensuring the stability of the decomposition. The above STL decomposition process is executed independently for each channel to obtain the trend component sequence, periodic component sequence, and residual component sequence for each channel. Then, the similar component sequences of all channels are combined to construct a multi-channel decomposition matrix, providing input for subsequent differential modeling.

[0041] This step designs three independent deep learning models for the three components obtained from STL decomposition: Trend, Seasonal, and Residual, which have different time scales and statistical properties. The three-model structure allows each model to learn a "normal pattern" in its specific subspace, thereby generating significant reconstruction errors when anomalies occur and achieving high-precision anomaly detection.

[0042] First, the trend sequence To characterize the long-term evolution of telemetry parameters, whose sequences exhibit significant time dependence, a multi-layer LSTM autoencoder structure is employed for the trend component, using the trend values ​​obtained from STL decomposition as input.

[0043] Let the trend window length be Construct training samples within each data segment

[0044] Then, it is standardized and input into the model. The input tensor is represented in terms of dimension as follows:

[0045] An LSTM encoder consists of two layers of recursive units connected in series. For the first layer of the encoder, its update equation is expressed as:

[0046]

[0047]

[0048]

[0049]

[0050]

[0051] in, Represents the hidden state of the trend model. This is the hidden dimension of the trend model.

[0052] The second LSTM layer receives the output sequence from the first layer and continues recursively encoding, ultimately using the hidden state of the last time step as the global representation of the trend sequence. The trend latent vector is then obtained through linear layer mapping.

[0053] in, This represents the potential spatial dimension of the trend. The decoder part uses the latent vector. As the initial input, by Repeatedly expanded to a length of sequence

[0054] A trend reconstruction sequence is generated step by step using a two-layer LSTM structure. The LSTM structure at the decoder end is the same as that at the encoder end, but the parameters are not shared. Finally, the output of each time step of the decoder is mapped to a scalar through a linear layer to obtain the trend reconstruction sequence. Its training objective is...

[0055] During training, batch gradient updates are performed on all trend segments using the Adam optimizer, and segment boundary constraints are employed to ensure that each sequence lies entirely within the same data segment. The LSTM autoencoder results are compared with those of the LSTM units as follows: Figure 4 As shown.

[0056] Periodic components It exhibits fixed period length and local waveform repetition characteristics, and since modeling it does not require capturing long-range dependencies, a one-dimensional convolutional autoencoder (1DCNN-AE) composed of multiple convolutional layers and upsampling structures is employed. The window length of the periodic sequence is determined by the number of period points obtained from STL decomposition. Decision, construct input

[0057] A convolutional encoder consists of multiple layers of one-dimensional convolutional units connected in series. For the encoder's... In terms of layers, there are

[0058] The kernel size for each convolutional layer is set to... Step size is The number of channels is The initial input is Max pooling can be used after the convolutional layer.

[0059] To further compress the periodic structure, after several layers of convolution and pooling, the sequence length is gradually reduced, and finally, at the end of the encoder, a flattening operation is performed to form a vector, which is then mapped to the periodic latent space.

[0060] in, This represents the latent spatial dimension of the period. The decoder recovers the sequence length through upsampling and deconvolution. Let the upsampling operation be...

[0061] And through deconvolution

[0062] Gradually restore the sequence to its length The final layer of the decoder uses a convolution with a kernel size of 1 to map the number of channels back to 1, resulting in a periodic reconstruction sequence.

[0063] Training objectives adopt

[0064] During training, it is also necessary to ensure that the samples in each periodic window are within the same data segment to avoid error propagation across segments.

[0065] Residual sequence Theoretically, it contains a noisy principal component, exhibiting a random distribution rather than temporal correlation, thus making it unsuitable for deep models such as autoencoders. The residual part is directly modeled using a statistical distribution approach, and parameter estimation is performed on the residual values ​​of all training segments. For the training segment set... Estimate the mean of all residuals.

[0066] With variance

[0067] The residual is expressed as the standardized deviation during the detection phase as follows:

[0068] Standardized bias characterizes the degree to which noise deviates from its normal distribution, serving as a quantitative indicator of anomalies in the residual direction. Since the residuals lack significant sequence structure, this method can obtain stable bias quantification results while preserving the noise properties of the residuals.

[0069] This step utilizes a multi-layer LSTM autoencoder structure based on the long-term evolution characteristics of the trend component. Within each data segment, training samples are constructed according to a set trend window length and standardized before being input into the model. The encoder performs global feature encoding on the trend sequence through two concatenated recursive units, mapping the hidden state of the last time step to a trend latent vector. The decoder uses this latent vector as its initial input and, after repeated expansion, gradually generates a trend reconstruction sequence through a two-layer LSTM structure with the same structure as the encoder but without parameter sharing. The training process employs batch gradient updates and the Adam optimizer, using segment boundary constraints to ensure that each sequence is completely within the same data segment, avoiding cross-segment error propagation. Based on the fixed period and local waveform repetition characteristics of the periodic component, a 1D-CNN autoencoder composed of multi-layer convolution and upsampling structures is used. The periodic sequence window length is determined based on the number of periodic points obtained from STL decomposition, and input samples are constructed. The encoder extracts periodic local structural features through multi-layer one-dimensional convolutional units in series, compresses them by max pooling, and then maps them to the periodic latent space through a flattening operation. The decoder gradually recovers the sequence length through upsampling and deconvolution, and finally maps the number of channels back to 1 through a convolution with a kernel size of 1 to obtain the periodic reconstructed sequence. Based on the characteristic that the residual components contain noise principal components and exhibit random distribution, a statistical distribution model is adopted. Statistical analysis is performed on the residual values ​​of all training segments to estimate the mean and variance parameters of the residuals. In the detection stage, the standardization deviation is calculated to characterize the degree of noise deviation from the normal distribution, serving as an anomaly quantification indicator in the residual direction, thus avoiding overfitting of the deep model to random noise.

[0070] Step 106: Calculate the reconstruction error of each model obtained from modeling, and perform weighted fusion of the trend component reconstruction error, periodic component reconstruction error and residual component deviation to obtain a comprehensive anomaly score. Construct a judgment threshold based on the normal sample distribution, and complete point-by-point anomaly judgment by combining the anomaly score and the threshold. At the same time, realize the anomaly source location according to the error weight distribution of each component.

[0071] After the trend model and cycle model are reconstructed respectively, the reconstruction error sequence of the trend direction can be obtained. Reconstruction error sequence with periodic direction Meanwhile, for the residual component, since the residual model uses a statistical distribution modeling approach, its bias can be given by the standardized residuals and expressed through absolute values. This indicates the degree of anomaly in the residual direction. The three components mentioned above reflect the deviation of the current moment from the normal trend structure, normal periodic structure, and normal noise distribution, respectively. To form a comprehensive score that can be used for anomaly identification on a unified time axis, the three error signals need to be fused, and a threshold needs to be estimated based on the data distribution under normal conditions to achieve the final point-by-point anomaly determination.

[0072] In time index At this point, the trend error, periodic error, and residual error are denoted as follows:

[0073] in, Derived from trend model Single-point reconstruction deviation, Derived from the periodic model Reconstruction deviation, This is the residual model pair Reconstruction deviation.

[0074] To obtain a uniform outlier score, the three errors are fused using a weighted linear method. Let the weighting coefficients for trend, period, and residual be denoted as follows: , and They satisfy

[0075] Fusion anomaly score is defined as

[0076] The fusion score The unified timeline reflects the combined degree of the three types of structural deviations and can be used in subsequent anomaly detection steps. This is to analyze the anomaly score sequence. To make effective discrimination, a threshold model needs to be constructed first based on the normal data from the training phase. Let the set of time indices corresponding to the training set be... Then, the abnormal subarray under normal conditions can be represented as:

[0077] Based on this, an adaptive threshold can be generated using a percentile-based statistical method. Let the percentile be... (For example or The anomaly threshold is defined based on the distribution of normal samples.

[0078] and satisfy

[0079] This threshold ensures that only a very small percentage of values ​​in normal data exceed it. This allows for the identification of potential outliers by observing deviations from the normal distribution. If the system has stable long-term operational experience, a fixed threshold can also be used directly. As an anomaly detection criterion, it is not necessary to rely on statistical percentiles.

[0080] S5.3 Point-by-point anomaly detection based on fusion anomaly scores After obtaining the fusion anomaly score With threshold Afterwards, each test time point can be evaluated individually. For each time point... If its abnormal score exceeds the threshold, i.e.

[0081] If the time point is not within a given time segment, it is considered abnormal; otherwise, it is considered normal. To avoid bias in judgment due to insufficient window size or incomplete structure at data segment boundaries, further boundary exclusion operations are required for time points near segment boundaries. If the time point is located within a certain data segment... The surrounding area, for example

[0082] Abnormal scores within this region will not be included in the final statistics, and their corresponding judgment labels can be set to an undefined state to avoid false positives.

[0083] After the above processing, the anomaly score sequence and threshold are combined to form the point-by-point anomaly output result of this application, thereby completing the multi-component anomaly detection process of satellite telemetry data based on a deep learning model. This step calculates the absolute values ​​of the trend component reconstruction error, periodic component reconstruction error, and residual component standardized deviation, respectively, and uses these three as the anomalous errors of the corresponding components. Weighting coefficients that are non-negative and sum to 1 are assigned to the three errors, and a comprehensive anomalous score is obtained through weighted linear summation. This score reflects the comprehensive degree of the three types of structural deviations on a unified time axis. Based on the set of anomalous scores corresponding to the normal data in the training set, a percentile statistical method is used to generate an adaptive threshold, ensuring that only a very small portion of the normal data values ​​exceed this threshold. If the system has stable long-term operating experience, a fixed threshold can also be directly used as the anomaly detection criterion. For all test time points, if the overall abnormal score exceeds the threshold, it is judged as abnormal; otherwise, it is considered normal. To avoid misjudgment due to insufficient window or incomplete structure at the data segment boundary, a boundary exclusion operation is performed on the time points in the area near the segment boundary. The abnormal score in this area is not included in the final statistics, and the judgment label is set to undefined. At the same time, according to the weight distribution of the three component errors, the source of the abnormality is located: trend abnormality corresponds to device aging, periodic abnormality corresponds to the disorder of operating rhythm, and residual abnormality corresponds to sudden changes in the external environment.

[0084] The aforementioned method for anomaly detection in spaceborne products based on multidimensional telemetry time-series STL decomposition first constructs a standardized time-series data table through multi-channel telemetry data loading and structured processing, laying a standardized data foundation for subsequent processing. The asynchronous multi-channel alignment and downsampling mechanism based on a fixed time window not only maps multidimensional data with different sampling frequencies and time offsets to a unified time axis, solving the time-series synchronization problem in multidimensional coupled analysis, but also utilizes the window mean to naturally suppress high-frequency random noise. Simultaneously, it handles irregular missing data through interpolation or pre- and post-filling, adapting to the characteristics of data segmentation. Secondly, the innovative STL time-series decomposition strategy decouples complex telemetry signals into three independent components: trend, period, and residual. This fundamentally avoids long-term, slowly changing trends being masked by short-term periodic fluctuations or noise, allowing subtle aging and degradation characteristics to be highlighted, significantly improving sensitivity to latent anomalies. Furthermore… This approach employs differentiated and independent modeling for the three components, taking into account their distinct characteristics. An LSTM autoencoder accurately captures the long-range dependence of the trend component, a 1D-CNN autoencoder effectively extracts the local structure of the periodic component, and statistical distribution modeling adapts to the noise characteristics of the residual component. This design avoids the forced fitting of deep models to Gaussian white noise when outlier samples are extremely scarce, thus mitigating the risk of overfitting and reducing the false alarm rate. Finally, a comprehensive outlier score is obtained by weighted fusion of the errors of the three components. An adaptive threshold is constructed by combining the distribution of normal samples, enabling accurate capture of all types of anomalies, including instantaneous outliers, progressive degradation, and waveform distortion. Furthermore, the error weight distribution of each component allows for intuitive location of the anomaly source, providing strong interpretability of the detection results and further ensuring the accuracy and reliability of anomaly detection. This approach comprehensively adapts to the complex characteristics of telemetry data from spaceborne products.

[0085] In one embodiment, the telemetry data of multi-channel spaceborne products is loaded and structured to construct a standardized time-series data table, including: The telemetry data of multi-channel spaceborne products are loaded and structured to construct a standardized time-series data table.

[0086] in, Indicates the first Each channel at time The telemetry value, A timestamp sequence, For satellite identification sequence, For the number of telemetry channels, This represents the number of valid sample points.

[0087] In one embodiment, asynchronous multi-channel alignment and downsampling of standardized time-series data based on a fixed time window includes: A unified window starting point is constructed based on the earliest timestamp of all channels. Continuous windows are generated with a fixed window length. For each channel, the sampling time is mapped to the corresponding window, and a set of data points within the window is constructed. The window mean is calculated based on the set of data points within the window. The window means of all channels are combined into a vector to obtain the multi-channel time series matrix after window alignment.

[0088] In one embodiment, generating a continuous window with a fixed window length is as follows:

[0089] in, Indicates a continuous window. Indicates the start of the window. Indicates a fixed window length. Indicates the valid sample point number; The set of data points within the window is constructed as follows:

[0090] in, Indicates channel The set of data points within the window. Indicates the sampling time.

[0091] In one embodiment, STL time series decomposition is performed on the multi-channel time series matrix to decouple the complex telemetry signal into three independent components: trend, period, and residual. First, the number of windows in each period is determined based on the known main period of the system and the time resolution after window alignment, and this is used as the seasonal window input for STL decomposition. Before decomposition, missing windows in the multi-channel time matrix are filled with linear interpolation and edge regions are filled to ensure signal continuity. The robust STL decomposition algorithm completes the decomposition through a three-stage iterative process based on LOESS local regression. The first stage performs LOESS smoothing fitting on the periodic window to extract the morphologically stable periodic components. The second stage applies large-window LOESS local regression to the deperiodic signal to capture trend components on the scale of multiple days to several weeks. The third stage obtains residual components containing sudden disturbances, noise and local anomalies through differential operation. During the iteration process, the interference of outliers on the fitting results of trend components and periodic components is suppressed by dynamically adjusting the weights. The above decomposition process is performed independently on each channel in the multi-channel time series matrix to obtain the trend component sequence, periodic component sequence and residual component sequence corresponding to each channel. Then, the similar component sequences of all channels are combined to construct a multi-channel decomposition matrix.

[0092] In one embodiment, differentiated independent modeling is performed for the three independent components: an LSTM autoencoder is used to model the trend component, a 1D-CNN autoencoder is used to model the periodic component, and a statistical distribution is used to model the residual component, including: The trend component is modeled using an LSTM autoencoder. Within each data segment, training samples are constructed according to the set trend window length and then standardized before being input into the model. The encoder encodes the global features of the trend sequence through two layers of recursive units connected in series. The decoder uses the encoded latent vector as the initial input and, after repeated expansion, gradually generates the trend reconstruction sequence through a two-layer LSTM structure that is consistent with the encoder structure but does not share parameters. The periodic components are modeled using a 1D-CNN autoencoder consisting of multi-layer convolution and upsampling structures. The periodic sequence window length is determined by the number of periodic points obtained from STL decomposition, and input samples are constructed. The encoder extracts periodic local structural features through multi-layer one-dimensional convolutional units in series. After compression by max pooling, the features are mapped to the periodic latent space through flattening. The decoder gradually recovers the sequence length through upsampling and deconvolution. Finally, the number of channels is mapped back to 1 through convolution with a kernel size of 1 to obtain the periodic reconstructed sequence. The residual components are modeled using a statistical distribution method. Statistical analysis is performed on the residual values ​​of all training segments to estimate the mean and variance parameters of the residuals. In the detection stage, the standardized deviation table is calculated as the residual reconstruction sequence.

[0093] In one embodiment, the encoder performs global feature encoding on the trend sequence through two layers of recursive units concatenated together, including: The update equation for the first recursive unit in the encoder is expressed as follows:

[0094]

[0095]

[0096]

[0097]

[0098]

[0099] in, Represents the hidden state of the trend model. As a hidden dimension of the trend model, , , , This represents the weight matrix corresponding to the input vector. Indicates the first tThe input vector at time t, Indicates the first The hidden state at any given moment; Indicates the first Cellular state at any given moment Represents the Gate of Oblivion Indicates the input gate. Indicates the output gate. Indicates the state of candidate cells. Indicates the first t The constantly updated state of the cell. , , , Indicates hidden state The corresponding weight matrix, , , , This indicates the corresponding bias term. This represents the Sigmoid activation function. This represents the hyperbolic tangent activation function. This indicates element-wise multiplication.

[0100] In one embodiment, the second recursive unit receives the output sequence of the first layer and continues recursively encoding, using the hidden state of the last time step as the global representation of the trend sequence, and obtaining the trend latent vector through linear layer mapping.

[0101] in, As a potential spatial dimension of trends, This represents the hidden state of the last layer of the LSTM at the last time step. Represents the weight matrix. This indicates the bias term.

[0102] In one embodiment, the residual components are modeled using a statistical distribution approach. Statistical analysis is performed on the residual values ​​of all training segments to estimate the mean and variance parameters of the residuals. During the detection phase, a standardized deviation table is calculated as the residual reconstruction sequence, including: The residual components are modeled using a statistical distribution approach. Statistical analysis is performed on the residual values ​​for all training segments to estimate the mean and variance parameters of the residuals.

[0103]

[0104] in, This represents the mean of the residuals. Represents the residual sequence. Represents the set of training segments. Indicates variance; During the detection phase, the standardized deviation table is calculated as the residual reconstruction sequence.

[0105] in, This represents the standard deviation of the residuals.

[0106] In one embodiment, a judgment threshold is constructed based on the normal sample distribution, and anomaly judgment is completed point-by-point by combining the anomaly score with the threshold. Simultaneously, the source of the anomaly is located based on the error weight distribution of each component, including: The judgment threshold is constructed based on the normal sample distribution.

[0107]

[0108] in, Represents an abnormal score sequence. This represents the set of time indices corresponding to the training set; Each test time point was evaluated individually. If the abnormal score exceeds the threshold, the time point is judged as abnormal; otherwise, it is considered normal.

[0109] In a specific embodiment, this embodiment constructs a simulation dataset based on the actual satellite orbital period and uses the detection framework based on STL decomposition and multi-model fusion proposed in this application for systematic testing and analysis.

[0110] The experimental data generated simulated data spanning one year based on the normal and degradation characteristics of real on-orbit telemetry data. The data is presented as multiple data segments of unequal length, fully reflecting the typical characteristics of segmented and irregularly missing data in spaceborne telemetry data, consistent with the difficulties described in the background technology. The data includes various typical anomaly patterns injected in the simulation, used to systematically evaluate the detection capability of this method: the experimental data simulates typical telemetry characteristics of medium Earth orbit (MEO) satellites. The basic sampling rate is 16 seconds, and the main period is set to 13 hours (simulated orbital period). The data generator includes complex periodic terms based on sine wave superposition, long-term trend terms in linear and logarithmic forms, and Gaussian white noise with a signal-to-noise ratio of approximately 30%. To comprehensively evaluate the model's capabilities, a test set containing the following characteristics was generated according to the code logic: Abnormal slope changes: an early signal of progressive degradation in the performance of simulated spaceborne products.

[0111] Outlier anomalies: Simulate sudden anomalies caused by transient interference or data jumps.

[0112] Overall drift anomaly: Simulates systematic deviations caused by device aging or environmental factors.

[0113] Combination anomalies: These are combinations of the three anomalies mentioned above, used to test the overall performance of a method in complex scenarios.

[0114] In this embodiment, based on the method flow of this application, the specific parameter configuration in the experiment is as follows: Preprocessing and Alignment: After removing extreme outliers using the MAD (Absolute Median Difference) algorithm, the raw data is downsampled and aligned in a 30-minute window to form a standard input sequence; STL decomposition parameters: Set the seasonal period window length for STL decomposition based on a 13-hour period and a 30-minute sampling interval. (Right now ); Model structure: The trend, periodic and residual models adopt LSTM autoencoder, 1D-CNN autoencoder and Gaussian distribution-based statistical model structure respectively. The sequence length is set to 26, the hidden layer dimension is 64, and the Adam optimizer is used for training with 30 iterations. Fusion and Judgment: The weighting for fusion of outlier scores is set as: trend term weighting. Periodic term weights Residual term weights Based on the distribution characteristics of the training set, the fixed judgment threshold in this embodiment is set to... .

[0115] Through a comprehensive evaluation of five test scenarios, the method in this application demonstrates the ability to decouple and identify anomalies at different time scales, as detailed below: like Figure 5 The diagram shows the original telemetry time series, downsampling time series, decomposed trend term time series, periodic term time series, residual term time series, anomaly score time series, and detection result time series for normal data (left column) and outlier data (right column). In the normal test set without injected anomalies, the fused anomaly score output by the model remains consistently low (mean much less than 1.0), and is always below the preset threshold of 6.0. This indicates that STL decomposition effectively isolates normal periodic fluctuations and trend changes, enabling the deep learning model to accurately reconstruct the "normal pattern," demonstrating that the method has an extremely low false alarm rate.

[0116] like Figure 5 As shown, experimental results demonstrate that outlier detection of the original sequence using statistical features yields correct results for simulated outlier anomalies. Because this method introduces MAD-based original hierarchical outlier labeling in the preprocessing stage, the model exhibits extremely high sensitivity to short-term, high-frequency mutations, enabling "instant detection."

[0117] like Figure 6 The diagram shows the original telemetry time series, downsampling time series, decomposed trend term time series, period term time series, residual term time series, anomaly score time series, and detection result time series of the overall drift data (left column) and slope change data (right column). When a step drift occurs, both the trend term and period term after STL decomposition are destroyed, leading to a simultaneous increase in the reconstruction error of the trend model and the period model. The fused anomaly score rapidly jumps to above 10.0 at the drift point (far exceeding the threshold of 6.0) and remains at a high alarm level during the subsequent continuous drift phase. This verifies that the method does not rely on a single threshold but rather identifies systematic biases based on "structural destruction."

[0118] like Figure 6 As shown, regarding the gradual slope change, experiments indicate that in the initial stage of anomalous injection (when the deviation is small), the anomalous score rises slowly; as the degradation deepens, the anomalous score steadily exceeds the threshold and exhibits a monotonically increasing trend. This characteristic is highly consistent with the aging curve of spaceborne products, proving that the trend component model successfully captures deviations from long-term evolutionary patterns.

[0119] like Figure 7 As shown, in complex scenarios involving early outliers, mid-term drift, and late-term slope changes, this method achieves full-process tracking. Experimental curves demonstrate that the anomaly score accurately depicts the three stages of fault evolution: early spike alarms, mid-term plateau high-level alarms, and late-term ramp-up alarms. In particular, the weight allocation strategy in the code effectively avoids mutual masking of different types of anomalies during the fusion process, achieving 100% coverage of composite fault modes.

[0120] To quantitatively evaluate the detection performance of this method, statistical validation was performed on a large-scale test set containing various injected anomalies. Experimental results show that the method achieves both high recall and high precision in anomaly detection tasks. Specifically, its recall rate for real anomalies exceeds 98.6%, demonstrating that the method can comprehensively capture the vast majority of anomaly patterns; simultaneously, its precision reaches approximately 97.2%, reflecting an extremely low false alarm rate and effectively ensuring the reliability of the early warning system. These two core indicators jointly confirm that, when dealing with complex telemetry data from spaceborne products, this application can achieve efficient and accurate identification of both latent and overt anomalies while nearly eliminating false alarms.

[0121] Beneficial effects of this application Compared with existing technologies, the spaceborne product anomaly detection method based on deep learning reconstruction error fusion of multidimensional telemetry time series STL decomposition provided in this application has the following significant advantages: (1) It breaks through the modeling bottleneck caused by multi-channel asynchronous sampling and strong noise interference, and is specifically suitable for multi-dimensional data fusion analysis in complex spaceborne environments. This application effectively maps multi-channel data with different sampling frequencies and different time offsets to a unified time axis through the "asynchronous alignment and downsampling based on fixed window" mechanism. This not only solves the time synchronization problem of multi-dimensional parameter coupling analysis, but also uses the window mean to naturally suppress high-frequency random noise in telemetry data, laying a high-quality data foundation for subsequent high-precision modeling.

[0122] (2) This application overcomes the shortcomings of traditional end-to-end deep learning models in extracting features at mixed time scales, achieving accurate capture of both long-term gradual trends and short-term periodic fluctuations. It innovatively introduces the "STL temporal decomposition" strategy, decoupling complex telemetry signals into three independent components: trend, period, and residual. This "divide and conquer" strategy fundamentally prevents the subtle long-term aging trends of spaceborne products from being masked by large-scale periodic fluctuations or noise, significantly improving the model's sensitivity to latent degradation features.

[0123] (3) By employing a differentiated "component-independent modeling" mechanism, the risk of overfitting is effectively avoided under conditions of extreme scarcity of abnormal samples, significantly reducing the false alarm rate. LSTM is used to capture long-range dependencies for trend terms, 1D-CNN is used to extract local structures for periodic terms, and a novel regression statistical distribution model is used for random residual terms. This design avoids the forced fitting of deep neural networks to Gaussian white noise, ensuring that the model only learns the true physical laws. Example data shows that this method, while maintaining a high recall rate (>98%), not only accurately identifies anomalies but also achieves an extremely high precision rate (>97%) by suppressing noise interference.

[0124] (4) A comprehensive judgment system for multi-dimensional reconstruction error fusion was constructed, which solves the limitation of a single model in being unable to take into account both sudden disturbances and gradual degradation. This application enables the detection system to have full-spectrum anomaly perception capability by weighted fusion of reconstruction errors in three subspaces: trend, period, and residual. Whether it is an instantaneous outlier, an early slope change, or waveform distortion, it can be effectively captured through a unified anomaly score, which is particularly suitable for identifying complex mixed anomalies.

[0125] (5) It provides interpretability superior to traditional black-box deep learning models, enabling ground personnel to quickly locate the source of anomalies. Since anomaly judgment is based on the fusion of errors from three independent components, when the system alarms, maintenance personnel can intuitively determine whether the anomaly originates from device aging (corresponding to trend anomalies), erratic operating rhythm (corresponding to periodic anomalies), or sudden changes in the external environment (corresponding to residual anomalies) based on the weight distribution of the component errors. This interpretability provides valuable engineering basis for fault location and health status assessment of spaceborne products.

[0126] In summary, this application effectively addresses the core challenges faced in on-orbit health monitoring of satellite-borne products, such as multidimensional asynchrony, trend masking, noise overfitting, and complex anomaly patterns, providing a robust and efficient technical means for achieving high-performance and highly reliable autonomous health management of satellites in orbit.

[0127] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed, and they can be performed in other orders. Furthermore, Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0128] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0129] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for detecting anomalies in spaceborne products based on multidimensional telemetry time-series STL decomposition, characterized in that, The method includes: A standardized time-series data table is constructed by loading and structuring telemetry data of multi-channel spaceborne products; asynchronous multi-channel alignment and downsampling are performed on the standardized time-series data based on a fixed time window, and multi-channel data with different sampling frequencies and different time offsets are mapped to a unified time axis to obtain a window-aligned multi-channel time-series matrix. The multi-channel time series matrix is ​​decomposed using STL to decouple the complex telemetry signal into three independent components: trend, period, and residual. Differential independent modeling is then performed on each of the three independent components: the trend component is modeled using an LSTM autoencoder, the period component is modeled using a 1D-CNN autoencoder, and the residual component is modeled using a statistical distribution. The reconstruction errors of each model obtained by computational modeling are weighted and fused to obtain a comprehensive anomaly score by combining the reconstruction errors of the trend component, the reconstruction errors of the periodic component, and the deviation of the residual component. A judgment threshold is constructed based on the normal sample distribution. The anomaly score and the threshold are combined to complete the point-by-point anomaly judgment. At the same time, the source of the anomaly is located according to the error weight distribution of each component.

2. The method according to claim 1, characterized in that, The telemetry data from multi-channel spaceborne products are loaded and structured to construct standardized time-series data tables, including: The telemetry data of multi-channel spaceborne products are loaded and structured to construct a standardized time-series data table. in, Indicates the first Each channel at time The telemetry value, A timestamp sequence, For satellite identification sequence, For the number of telemetry channels, This represents the number of valid sample points.

3. The method according to claim 1, characterized in that, Asynchronous multi-channel alignment and downsampling of the standardized time-series data based on a fixed time window include: A unified window starting point is constructed based on the earliest timestamp of all channels. Continuous windows are generated with a fixed window length. For each channel, the sampling time is mapped to the corresponding window, and a set of data points within the window is constructed. The window mean is calculated based on the set of data points within the window. The window means of all channels are combined into a vector to obtain a multi-channel time series matrix after window alignment.

4. The method according to claim 3, characterized in that, Generating continuous windows with a fixed window length is as follows: in, Indicates a continuous window. Indicates the start of the window. Indicates a fixed window length. Indicates the valid sample point number; The set of data points within the window is constructed as follows: in, Indicates channel The set of data points within the window. Indicates the sampling time.

5. The method according to claim 1, characterized in that, The multi-channel time series matrix is ​​decomposed using STL time series decomposition to decouple the complex telemetry signal into three independent components: trend, period, and residual. First, the number of windows in each period is determined based on the known main period of the system and the time resolution after window alignment, and this is used as the seasonal window input for STL decomposition. Before decomposition, missing windows in the multi-channel time matrix are filled with linear interpolation and edge regions are filled to ensure signal continuity. A robust STL decomposition algorithm is adopted, which completes the decomposition through a three-stage iterative process based on LOESS local regression. The first stage performs LOESS smoothing fitting on the periodic window to extract the morphologically stable periodic components. The second stage applies large-window LOESS local regression to the deperiodic signal to capture trend components on the scale of multiple days to several weeks. The third stage obtains residual components containing sudden disturbances, noise and local anomalies through difference operations. During the iteration process, the weights are dynamically adjusted to suppress the interference of outliers on the fitting results of trend components and periodic components. The above decomposition process is performed independently on each channel in the multi-channel time series matrix to obtain the trend component sequence, periodic component sequence and residual component sequence corresponding to each channel. Then, the similar component sequences of all channels are combined to construct a multi-channel decomposition matrix.

6. The method according to claim 1, characterized in that, Differential independent modeling is performed for the three independent components. The trend component is modeled using an LSTM autoencoder, the periodic component using a 1D-CNN autoencoder, and the residual component using a statistical distribution model, including: The trend component is modeled using an LSTM autoencoder. Within each data segment, training samples are constructed according to the set trend window length and then standardized before being input into the model. The encoder encodes the global features of the trend sequence through two layers of recursive units connected in series. The decoder uses the encoded latent vector as the initial input and, after repeated expansion, gradually generates the trend reconstruction sequence through a two-layer LSTM structure that is consistent with the encoder structure but does not share parameters. The periodic components are modeled using a 1D-CNN autoencoder consisting of multi-layer convolution and upsampling structures. The periodic sequence window length is determined by the number of periodic points obtained from STL decomposition, and input samples are constructed. The encoder extracts periodic local structural features through multi-layer one-dimensional convolutional units in series. After compression by max pooling, the features are mapped to the periodic latent space through flattening. The decoder gradually recovers the sequence length through upsampling and deconvolution. Finally, the number of channels is mapped back to 1 through convolution with a kernel size of 1 to obtain the periodic reconstructed sequence. The residual components are modeled using a statistical distribution method. Statistical analysis is performed on the residual values ​​of all training segments to estimate the mean and variance parameters of the residuals. In the detection stage, the standardized deviation table is calculated as the residual reconstruction sequence.

7. The method according to claim 6, characterized in that, The encoder performs global feature encoding on the trend sequence through two layers of recursive units concatenated together, including: The update equation for the first recursive unit in the encoder is expressed as follows: in, Represents the hidden state of the trend model. As a hidden dimension of the trend model, , , , This represents the weight matrix corresponding to the input vector. Indicates the first t The input vector at time t, Indicates the first The hidden state at any given moment; Indicates the first Cellular state at any given moment Represents the Gate of Oblivion Indicates the input gate. Indicates the output gate. Indicates the state of candidate cells. Indicates the first t The constantly updated state of the cell. , , , Indicates hidden state The corresponding weight matrix, , , , This indicates the corresponding bias term. This represents the Sigmoid activation function. This represents the hyperbolic tangent activation function. This indicates element-wise multiplication.

8. The method according to claim 7, characterized in that, The method further includes: The second recursive unit receives the output sequence from the first layer and continues recursively encoding. It uses the hidden state of the last time step as the global representation of the trend sequence, and after mapping through a linear layer, obtains the trend latent vector. in, As a potential spatial dimension of trends, This represents the hidden state of the last layer of the LSTM at the last time step. Represents the weight matrix. This indicates the bias term.

9. The method according to claim 1, characterized in that, The residual components are modeled using a statistical distribution approach. Statistical analysis is performed on the residual values ​​of all training segments to estimate the mean and variance parameters of the residuals. During the detection phase, a standardized deviation table is calculated as the residual reconstruction sequence, including: The residual components are modeled using a statistical distribution approach. Statistical analysis is performed on the residual values ​​for all training segments to estimate the mean and variance parameters of the residuals. in, This represents the mean of the residuals. Represents the residual sequence. Represents the set of training segments. Indicates variance; During the detection phase, the standardized deviation table is calculated as the residual reconstruction sequence. in, This represents the standard deviation of the residuals.

10. The method according to claim 1, characterized in that, A judgment threshold is constructed based on the normal sample distribution. Anomaly judgment is completed point-by-point by combining the anomaly score with the threshold. Simultaneously, the source of the anomaly is located based on the error weight distribution of each component, including: The judgment threshold is constructed based on the normal sample distribution. in, Represents an abnormal score sequence. This represents the set of time indices corresponding to the training set; Each test time point was evaluated individually. If the abnormal score exceeds the threshold, the time point is judged as abnormal; otherwise, it is considered normal.

Images