CN-121978722-A - Satellite-borne product anomaly detection method based on multidimensional telemetry time sequence STL decomposition

CN121978722ACN 121978722 ACN121978722 ACN 121978722ACN-121978722-A

Abstract

The application relates to a satellite-borne product anomaly detection method based on multi-dimensional telemetry time sequence STL decomposition. The method comprises the steps of constructing a standardized time sequence data table, completing asynchronous multichannel alignment and downsampling through a fixed time window, mapping data with different sampling frequencies and time offsets to a unified time axis, and generating a multichannel time sequence matrix. And then performing STL time sequence decomposition on the matrix, decoupling three independent components of trend, period and residual error, and performing differential modeling by adopting an LSTM self-encoder, a 1D-CNN self-encoder and statistical distribution respectively. And finally, calculating reconstruction errors of the models, weighting and fusing to obtain comprehensive anomaly scores, finishing point-by-point anomaly judgment based on normal sample distribution construction threshold values, positioning anomaly sources according to the error weight distribution of each component, efficiently solving the problems of multi-channel asynchronism, strong noise and the like, and realizing high-precision anomaly detection.

Inventors

XIANG YAN
LI JIAYU
PEI LING

Assignees

上海交通大学

Dates

Publication Date: 20260505
Application Date: 20260113

Claims (10)

1. A satellite-borne product anomaly detection method based on multi-dimensional telemetry time sequence STL decomposition is characterized by comprising the following steps: carrying out asynchronous multichannel alignment and downsampling on the standardized time sequence data based on a fixed time window, and mapping multichannel data with different sampling frequencies and different time offsets to a unified time axis to obtain a multichannel time sequence matrix after window alignment; Performing STL time sequence decomposition on the multichannel time sequence matrix, and decoupling a complex telemetry signal into three independent components of trend, period and residual error; differential independent modeling is respectively carried out on the three independent components, LSTM self-encoder modeling is adopted for trend components, 1D-CNN self-encoder modeling is adopted for periodic components, and statistical distribution modeling is adopted for residual components; calculating reconstruction errors of the models obtained through modeling, carrying out weighted fusion on the trend component reconstruction errors, the periodic component reconstruction errors and the residual component deviation to obtain comprehensive anomaly scores, constructing a judgment threshold value based on normal sample distribution, combining the anomaly scores and the threshold value to finish point-by-point anomaly judgment, and simultaneously realizing anomaly source positioning according to the component error weight distribution.
2. The method of claim 1, wherein loading and structuring the multi-channel on-board product telemetry data to construct a standardized time series data table comprises: Loading and structuring the telemetry data of the multichannel satellite-borne product to construct a standardized time sequence data table as Wherein, the Represent the first The individual channels are at the moment Is used for the remote measurement of the value of (c), In order to be a sequence of time stamps, For the identification sequence of the satellite(s), For the number of telemetry channels, Points for valid samples.
3. The method of claim 1, wherein asynchronously multi-channel aligning and downsampling the normalized timing data based on a fixed time window comprises: And constructing a unified window starting point according to the earliest time stamp of all channels, generating a continuous window with a fixed window length, mapping the sampling time to a corresponding window for each channel, constructing a data point set in the window, calculating a window mean value according to the data point set in the window, forming a vector by the window mean value of all channels, and obtaining a multi-channel time sequence matrix after window alignment.
4. A method according to claim 3, wherein generating successive windows at a fixed window length is: Wherein, the A continuous window is represented and a continuous window is represented, Indicating the start point of the window, Indicating that the length of the window is fixed, A valid sample point sequence number is represented; the set of data points within the building window are: Wherein, the Representation channel Is a set of data points within a window of (a), Indicating the sampling instant.
5. The method of claim 1, wherein STL timing decomposition of the multi-channel timing matrix decouples complex telemetry signals into three independent components of trend, period, and residual, comprising: determining the number of windows contained in each period based on the time resolution of the system after the known main period is aligned with the windows, and taking the window number as seasonal window input of STL decomposition; The STL decomposition algorithm with a robust mode is adopted, the decomposition is completed through a three-stage iteration process based on local regression of the LOESS, the LOESS smooth fitting is carried out on a periodic window in the first stage to extract a periodic component with stable morphology, the local regression of the large window LOESS is applied to the signal after the decyclization in the second stage to capture trend components with the scale of multiple days to several weeks, and the residual components containing sudden disturbance, noise and local abnormality are obtained through differential operation in the third stage; And independently executing the decomposition flow for each channel in the multi-channel time sequence matrix, respectively acquiring a trend component sequence, a periodic component sequence and a residual component sequence corresponding to each channel, and then combining the similar component sequences of all the channels to construct the multi-channel decomposition matrix.
6. The method of claim 1, wherein the differential independent modeling is performed for each of the three independent components, wherein the trend component is modeled using an LSTM self-encoder, wherein the periodic component is modeled using a 1D-CNN self-encoder, wherein the residual component is modeled using a statistical distribution, comprising: Modeling the trend component by adopting an LSTM self-encoder, constructing a training sample in each data segment according to the set trend window length, standardizing the training sample, inputting the standardized training sample into a model, carrying out global feature coding on the trend sequence by the encoder through two layers of recursion units in series, taking a potential vector obtained by coding as initial input by the decoder, and gradually generating a trend reconstruction sequence through two layers of LSTM structures which are consistent with the structure of the encoder but have unshared parameters after repeated expansion; Modeling a 1D-CNN self-encoder formed by a multi-layer convolution and an up-sampling structure for the periodic component, determining the window length of a periodic sequence according to the number of periodic points obtained by STL decomposition, constructing an input sample, extracting the local structural characteristics of the periodic sequence in series by using a multi-layer one-dimensional convolution unit, mapping the local structural characteristics to a periodic potential space through flattening operation after maximum pooling compression, gradually recovering the sequence length by up-sampling and deconvolution, and finally mapping the number of channels back to 1 through convolution with a kernel size of 1 to obtain a periodic reconstruction sequence; Modeling the residual components in a statistical distribution mode, carrying out statistical analysis on residual values of all training segments, estimating mean and variance parameters of residual, and using a calculated standard deviation table as a residual reconstruction sequence in a detection stage.
7. The method of claim 6, wherein the encoder globally feature codes the trend sequence by concatenating two layers of recursive units, comprising: the update equation of the first layer recursion unit in the encoder is expressed as Wherein, the Represents the hidden state of the trend model, As a hidden dimension of the trend model, , , , Representing the weight matrix to which the input vector corresponds, The input vector at time t is represented, Represent the first A hidden state from time to time; represent the first The state of the cell at the moment in time, Indicating that the door is left to be forgotten, The input gate is shown as being provided with a display, The output gate is shown as being provided with a display, Indicating the state of the candidate cell, Indicating the updated cell state at time t, , , , Indicating hidden state The corresponding weight matrix is used to determine the weight matrix, , , , Representing the corresponding bias term(s), Representing the Sigmoid activation function, Representing the hyperbolic tangent activation function, Representing a per-element multiplication operation.
8. The method of claim 7, wherein the method further comprises: The second layer recursion unit receives the output sequence of the first layer and continues recursion coding, uses the hidden state of the last time step as the global representation of the trend sequence, and obtains a trend potential vector by linear layer mapping as follows Wherein, the As a potential spatial dimension of the trend, Indicating the hidden state of the last layer of LSTM at the last time step, The weight matrix is represented by a matrix of weights, Representing the bias term.
9. The method according to claim 1, wherein modeling the residual components in a statistical distribution manner, performing a statistical analysis on residual values of all training segments, estimating a mean and variance parameter of the residual, and reconstructing the sequence by calculating a standardized deviation table as the residual in the detection stage, comprises: Modeling residual components in a statistical distribution mode, carrying out statistical analysis on residual values of all training segments, wherein the mean value and variance parameters of estimated residual are respectively Wherein, the Represents the mean value of the residual error, A sequence of residuals is represented and, A set of training segments is represented and, Representing the variance; Reconstructing the sequence as a residual by computing a normalized deviation table during the detection phase Wherein, the Representing the residual standard deviation.
10. The method of claim 1, wherein constructing a decision threshold based on normal sample distribution, combining anomaly scores with the threshold to complete point-by-point anomaly decision, and simultaneously implementing anomaly source localization according to component error weight distribution, comprises: constructing a decision threshold based on normal sample distribution as Wherein, the Representing a sequence of anomaly scores, Representing a corresponding time index set in the training set; Point-by-point judgment is carried out on all the test time points, and the time points are If the abnormality score exceeds the threshold, the time point is judged to be abnormal, otherwise, the time point is judged to be normal.

Description

Satellite-borne product anomaly detection method based on multidimensional telemetry time sequence STL decomposition Technical Field The application relates to the technical field of spacecraft health management and satellite on-orbit intelligent monitoring, in particular to a satellite-borne product anomaly detection method based on multi-dimensional telemetry time sequence STL decomposition. Background The on-board product is used as a core component of the spacecraft, and the on-orbit running state of the on-board product directly determines success or failure of the spaceflight task. The anomaly detection of the satellite-borne product is mainly realized by means of telemetry data, and the fault precursors and the anomaly states of the equipment are identified by analyzing multichannel time sequence parameters such as voltage, current, temperature and the like. However, the prior art faces a plurality of bottlenecks in practical application, and is difficult to meet the high-precision detection requirement of a complex satellite-borne environment: Firstly, satellite-borne telemetry data has the problems of multi-channel asynchronous sampling and strong noise interference. The sampling frequency of different sensors has larger difference, time offset exists, so that multidimensional data is difficult to directly carry out coupling analysis, and meanwhile, a large amount of high-frequency random noise is introduced by factors such as space electromagnetic radiation, inherent errors of the sensors and the like, so that the data quality is seriously influenced. The traditional method mostly adopts simple interpolation alignment or direct splicing of data, so that the problem of timing synchronization cannot be solved, noise is difficult to effectively inhibit, and the modeling basis is weak. Secondly, the satellite-borne telemetry signal comprises mixed time scale features such as long-term slowly-varying trend, short-term periodic fluctuation, random residual error and the like, and the traditional end-to-end deep learning model (such as single LSTM and CNN) adopts a one-tool modeling mode, so that the characteristics of different scales are difficult to capture simultaneously. The tiny long-term aging trend is often covered by large-amplitude periodic fluctuation or noise, so that the sensitivity of the model to recessive degradation characteristics is insufficient, and early fault early warning cannot be realized. Furthermore, abnormal samples of the satellite-borne product are extremely scarce, most of the operation time periods are in a normal state, the traditional deep neural network is easy to forcedly fit Gaussian white noise in the training process, so that the risk of overfitting is increased, and the problem of high false alarm rate is solved. The existing method is relieved by data enhancement or regularization means, but an important unit for providing each function support in the satellite system is not distinguished from a modeling mechanism, and the long-term stability of the method directly influences the overall performance of the satellite. Under a complex space environment, the satellite-borne product can be influenced by various factors such as radiation environment change, thermal control condition fluctuation, device aging and the like, so that gradual degradation of performances such as output parameter drift, abnormality and the like can occur. The traditional on-orbit health monitoring mode is generally based on fixed threshold judgment of a single telemetry index, and is difficult to identify a recessive performance degradation trend and capture a relevance abnormal mode existing among multiple parameters. At present, the following technical difficulties mainly exist in the health monitoring of on-orbit running states of satellite-borne products: (1) Multi-channel asynchronous sampling. Different measurement links have different sampling periods and sampling offsets, and the time stamps of the multiple channels are different. Because the control system is different from the clock synchronization mechanism of the payload, the multi-channel data is difficult to directly model uniformly. (2) Abnormal samples are extremely rare. On-orbit fault cases of the satellite-borne products are few, and few abnormal data are marked for supervised learning. The traditional classification abnormality detection method based on supervised learning is difficult to be applied. (3) The long-term slow-change trend is difficult to capture. Aging and performance changes in satellite-borne products are often manifested as slow drift lasting months or even years. The short time window can not effectively sense the long-term trend, and the long window can cause the problems of overlarge calculation cost, long training time, insufficient instantaneity and the like. (4) The multidimensional parameter coupling relation is complex. On-board products contain a number of key telemetry parameters that exist in co